LCOV - code coverage report
Current view: top level - lib/env_dpdk - memory.c (source / functions) Hit Total Coverage
Test: ut_cov_unit.info Lines: 0 593 0.0 %
Date: 2024-12-09 14:01:16 Functions: 0 34 0.0 %

          Line data    Source code
       1             : /*   SPDX-License-Identifier: BSD-3-Clause
       2             :  *   Copyright (C) 2017 Intel Corporation.
       3             :  *   All rights reserved.
       4             :  */
       5             : 
       6             : #include "spdk/stdinc.h"
       7             : 
       8             : #include "env_internal.h"
       9             : #include "pci_dpdk.h"
      10             : 
      11             : #include <rte_config.h>
      12             : #include <rte_memory.h>
      13             : #include <rte_eal_memconfig.h>
      14             : #include <rte_dev.h>
      15             : #include <rte_pci.h>
      16             : 
      17             : #include "spdk_internal/assert.h"
      18             : 
      19             : #include "spdk/assert.h"
      20             : #include "spdk/likely.h"
      21             : #include "spdk/queue.h"
      22             : #include "spdk/util.h"
      23             : #include "spdk/memory.h"
      24             : #include "spdk/env_dpdk.h"
      25             : #include "spdk/log.h"
      26             : 
      27             : #ifdef __linux__
      28             : #include <linux/version.h>
      29             : #if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 6, 0)
      30             : #include <linux/vfio.h>
      31             : #include <rte_vfio.h>
      32             : 
      33             : struct spdk_vfio_dma_map {
      34             :         struct vfio_iommu_type1_dma_map map;
      35             :         TAILQ_ENTRY(spdk_vfio_dma_map) tailq;
      36             : };
      37             : 
      38             : struct vfio_cfg {
      39             :         int fd;
      40             :         bool enabled;
      41             :         bool noiommu_enabled;
      42             :         unsigned device_ref;
      43             :         TAILQ_HEAD(, spdk_vfio_dma_map) maps;
      44             :         pthread_mutex_t mutex;
      45             : };
      46             : 
      47             : static struct vfio_cfg g_vfio = {
      48             :         .fd = -1,
      49             :         .enabled = false,
      50             :         .noiommu_enabled = false,
      51             :         .device_ref = 0,
      52             :         .maps = TAILQ_HEAD_INITIALIZER(g_vfio.maps),
      53             :         .mutex = PTHREAD_MUTEX_INITIALIZER
      54             : };
      55             : #endif
      56             : #endif
      57             : 
      58             : #if DEBUG
      59             : #define DEBUG_PRINT(...) SPDK_ERRLOG(__VA_ARGS__)
      60             : #else
      61             : #define DEBUG_PRINT(...)
      62             : #endif
      63             : 
      64             : struct map_page_cfg {
      65             :         uint64_t shift;
      66             :         uint64_t size;
      67             :         uint64_t mask;
      68             :         uint64_t num_pages_per_gb;
      69             : };
      70             : 
      71             : /**
      72             :  * g_map_page_cfg can not be static because it is used in the inline function
      73             :  * spdk_mem_map_translate.
      74             :  */
      75             : struct map_page_cfg g_map_page_cfg = {
      76             :         .shift = SHIFT_2MB,
      77             :         .size = VALUE_2MB,
      78             :         .mask = MASK_2MB,
      79             :         .num_pages_per_gb = 1UL << (SHIFT_1GB - SHIFT_2MB),
      80             : };
      81             : 
      82             : #define MAP_PAGE_SHIFT                  (g_map_page_cfg.shift)
      83             : #define MAP_PAGE_SIZE                   (g_map_page_cfg.size)
      84             : #define MAP_PAGE_MASK                   (g_map_page_cfg.mask)
      85             : #define MAP_NUM_PAGES_PER_GB    (g_map_page_cfg.num_pages_per_gb)
      86             : 
      87             : #define MAP_256TB_IDX(vfn_page) ((vfn_page) >> (SHIFT_1GB - MAP_PAGE_SHIFT))
      88             : #define MAP_1GB_IDX(vfn_page)   ((vfn_page) & ((1ULL << (SHIFT_1GB - MAP_PAGE_SHIFT)) - 1))
      89             : #define MAP_PAGE_OFFSET(ptr)    (((uintptr_t)(ptr)) & MAP_PAGE_MASK)
      90             : 
      91             : /* Page is registered */
      92             : #define REG_MAP_REGISTERED      (1ULL << 62)
      93             : 
      94             : /* A notification region barrier. The page translation entry that's marked
      95             :  * with this flag must be unregistered separately. This allows contiguous
      96             :  * regions to be unregistered in the same chunks they were registered.
      97             :  */
      98             : #define REG_MAP_NOTIFY_START    (1ULL << 63)
      99             : 
     100             : /* Translation of a single page. */
     101             : struct map_page {
     102             :         uint64_t translation;
     103             : };
     104             : 
     105             : /* Second-level map table indexed by bits [page_shift..29] of the virtual address.
     106             :  * Each entry contains the address translation or error for entries that haven't
     107             :  * been retrieved yet.
     108             :  */
     109             : struct map_1gb {
     110             :         struct map_page map[0];
     111             :         /**
     112             :          * Page table space.
     113             :          * Do not put any fields after this!
     114             :          */
     115             : };
     116             : 
     117             : #define MAP_SIZE_OF_MAP_1GB (sizeof(struct map_1gb) + MAP_NUM_PAGES_PER_GB * sizeof(struct map_page))
     118             : 
     119             : /* Top-level map table indexed by bits [30..47] of the virtual address.
     120             :  * Each entry points to a second-level map table or NULL.
     121             :  */
     122             : struct map_256tb {
     123             :         struct map_1gb *map[1ULL << (SHIFT_256TB - SHIFT_1GB)];
     124             : };
     125             : 
     126             : /* Page-granularity memory address translation */
     127             : struct spdk_mem_map {
     128             :         struct map_256tb map_256tb;
     129             :         pthread_mutex_t mutex;
     130             :         uint64_t default_translation;
     131             :         struct spdk_mem_map_ops ops;
     132             :         void *cb_ctx;
     133             :         TAILQ_ENTRY(spdk_mem_map) tailq;
     134             : };
     135             : 
     136             : /* Registrations map. The 64 bit translations are bit fields with the
     137             :  * following layout (starting with the low bits):
     138             :  *    0 - 61 : reserved
     139             :  *   62 - 63 : flags
     140             :  */
     141             : static struct spdk_mem_map *g_mem_reg_map;
     142             : static TAILQ_HEAD(spdk_mem_map_head, spdk_mem_map) g_spdk_mem_maps =
     143             :         TAILQ_HEAD_INITIALIZER(g_spdk_mem_maps);
     144             : static pthread_mutex_t g_spdk_mem_map_mutex = PTHREAD_MUTEX_INITIALIZER;
     145             : 
     146             : static bool g_legacy_mem;
     147             : static bool g_huge_pages = true;
     148             : 
     149             : /*
     150             :  * Walk the currently registered memory via the main memory registration map
     151             :  * and call the new map's notify callback for each virtually contiguous region.
     152             :  */
     153             : static int
     154           0 : mem_map_notify_walk(struct spdk_mem_map *map, enum spdk_mem_map_notify_action action)
     155             : {
     156             :         size_t idx_256tb;
     157             :         uint64_t idx_1gb;
     158           0 :         uint64_t contig_start = UINT64_MAX;
     159           0 :         uint64_t contig_end = UINT64_MAX;
     160             :         struct map_1gb *map_1gb;
     161             :         int rc;
     162             : 
     163           0 :         if (!g_mem_reg_map) {
     164           0 :                 return -EINVAL;
     165             :         }
     166             : 
     167             :         /* Hold the memory registration map mutex so no new registrations can be added while we are looping. */
     168           0 :         pthread_mutex_lock(&g_mem_reg_map->mutex);
     169             : 
     170           0 :         for (idx_256tb = 0;
     171           0 :              idx_256tb < sizeof(g_mem_reg_map->map_256tb.map) / sizeof(g_mem_reg_map->map_256tb.map[0]);
     172           0 :              idx_256tb++) {
     173           0 :                 map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
     174             : 
     175           0 :                 if (!map_1gb) {
     176           0 :                         if (contig_start != UINT64_MAX) {
     177             :                                 /* End of of a virtually contiguous range */
     178           0 :                                 rc = map->ops.notify_cb(map->cb_ctx, map, action,
     179           0 :                                                         (void *)contig_start,
     180           0 :                                                         contig_end - contig_start + MAP_PAGE_SIZE);
     181             :                                 /* Don't bother handling unregister failures. It can't be any worse */
     182           0 :                                 if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
     183           0 :                                         goto err_unregister;
     184             :                                 }
     185           0 :                         }
     186           0 :                         contig_start = UINT64_MAX;
     187           0 :                         continue;
     188             :                 }
     189             : 
     190           0 :                 for (idx_1gb = 0; idx_1gb < MAP_NUM_PAGES_PER_GB; idx_1gb++) {
     191           0 :                         if ((map_1gb->map[idx_1gb].translation & REG_MAP_REGISTERED) &&
     192           0 :                             (contig_start == UINT64_MAX ||
     193           0 :                              (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) == 0)) {
     194             :                                 /* Rebuild the virtual address from the indexes */
     195           0 :                                 uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << MAP_PAGE_SHIFT);
     196             : 
     197           0 :                                 if (contig_start == UINT64_MAX) {
     198           0 :                                         contig_start = vaddr;
     199           0 :                                 }
     200             : 
     201           0 :                                 contig_end = vaddr;
     202           0 :                         } else {
     203           0 :                                 if (contig_start != UINT64_MAX) {
     204             :                                         /* End of of a virtually contiguous range */
     205           0 :                                         rc = map->ops.notify_cb(map->cb_ctx, map, action,
     206           0 :                                                                 (void *)contig_start,
     207           0 :                                                                 contig_end - contig_start + MAP_PAGE_SIZE);
     208             :                                         /* Don't bother handling unregister failures. It can't be any worse */
     209           0 :                                         if (rc != 0 && action == SPDK_MEM_MAP_NOTIFY_REGISTER) {
     210           0 :                                                 goto err_unregister;
     211             :                                         }
     212             : 
     213             :                                         /* This page might be a part of a neighbour region, so process
     214             :                                          * it again. The idx_1gb will be incremented immediately.
     215             :                                          */
     216           0 :                                         idx_1gb--;
     217           0 :                                 }
     218           0 :                                 contig_start = UINT64_MAX;
     219             :                         }
     220           0 :                 }
     221           0 :         }
     222             : 
     223           0 :         pthread_mutex_unlock(&g_mem_reg_map->mutex);
     224           0 :         return 0;
     225             : 
     226             : err_unregister:
     227             :         /* Unwind to the first empty translation so we don't unregister
     228             :          * a region that just failed to register.
     229             :          */
     230           0 :         idx_256tb = MAP_256TB_IDX((contig_start >> MAP_PAGE_SHIFT) - 1);
     231           0 :         idx_1gb = MAP_1GB_IDX((contig_start >> MAP_PAGE_SHIFT) - 1);
     232           0 :         contig_start = UINT64_MAX;
     233           0 :         contig_end = UINT64_MAX;
     234             : 
     235             :         /* Unregister any memory we managed to register before the failure */
     236           0 :         for (; idx_256tb < SIZE_MAX; idx_256tb--) {
     237           0 :                 map_1gb = g_mem_reg_map->map_256tb.map[idx_256tb];
     238             : 
     239           0 :                 if (!map_1gb) {
     240           0 :                         if (contig_end != UINT64_MAX) {
     241             :                                 /* End of of a virtually contiguous range */
     242           0 :                                 map->ops.notify_cb(map->cb_ctx, map,
     243             :                                                    SPDK_MEM_MAP_NOTIFY_UNREGISTER,
     244           0 :                                                    (void *)contig_start,
     245           0 :                                                    contig_end - contig_start + MAP_PAGE_SIZE);
     246           0 :                         }
     247           0 :                         contig_end = UINT64_MAX;
     248           0 :                         continue;
     249             :                 }
     250             : 
     251           0 :                 for (; idx_1gb < UINT64_MAX; idx_1gb--) {
     252             :                         /* Rebuild the virtual address from the indexes */
     253           0 :                         uint64_t vaddr = (idx_256tb << SHIFT_1GB) | (idx_1gb << MAP_PAGE_SHIFT);
     254           0 :                         if ((map_1gb->map[idx_1gb].translation & REG_MAP_REGISTERED) &&
     255           0 :                             (contig_end == UINT64_MAX || (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) == 0)) {
     256             : 
     257           0 :                                 if (contig_end == UINT64_MAX) {
     258           0 :                                         contig_end = vaddr;
     259           0 :                                 }
     260           0 :                                 contig_start = vaddr;
     261           0 :                         } else {
     262           0 :                                 if (contig_end != UINT64_MAX) {
     263           0 :                                         if (map_1gb->map[idx_1gb].translation & REG_MAP_NOTIFY_START) {
     264           0 :                                                 contig_start = vaddr;
     265           0 :                                         }
     266             :                                         /* End of of a virtually contiguous range */
     267           0 :                                         map->ops.notify_cb(map->cb_ctx, map,
     268             :                                                            SPDK_MEM_MAP_NOTIFY_UNREGISTER,
     269           0 :                                                            (void *)contig_start,
     270           0 :                                                            contig_end - contig_start + MAP_PAGE_SIZE);
     271           0 :                                 }
     272           0 :                                 contig_end = UINT64_MAX;
     273             :                         }
     274           0 :                 }
     275           0 :                 idx_1gb = MAP_NUM_PAGES_PER_GB - 1;
     276           0 :         }
     277             : 
     278           0 :         pthread_mutex_unlock(&g_mem_reg_map->mutex);
     279           0 :         return rc;
     280           0 : }
     281             : 
     282             : struct spdk_mem_map *
     283           0 : spdk_mem_map_alloc(uint64_t default_translation, const struct spdk_mem_map_ops *ops, void *cb_ctx)
     284             : {
     285             :         struct spdk_mem_map *map;
     286             :         int rc;
     287             :         size_t i;
     288             : 
     289           0 :         map = calloc(1, sizeof(*map));
     290           0 :         if (map == NULL) {
     291           0 :                 return NULL;
     292             :         }
     293             : 
     294           0 :         if (pthread_mutex_init(&map->mutex, NULL)) {
     295           0 :                 free(map);
     296           0 :                 return NULL;
     297             :         }
     298             : 
     299           0 :         map->default_translation = default_translation;
     300           0 :         map->cb_ctx = cb_ctx;
     301           0 :         if (ops) {
     302           0 :                 map->ops = *ops;
     303           0 :         }
     304             : 
     305           0 :         if (ops && ops->notify_cb) {
     306           0 :                 pthread_mutex_lock(&g_spdk_mem_map_mutex);
     307           0 :                 rc = mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_REGISTER);
     308           0 :                 if (rc != 0) {
     309           0 :                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     310           0 :                         DEBUG_PRINT("Initial mem_map notify failed\n");
     311           0 :                         pthread_mutex_destroy(&map->mutex);
     312           0 :                         for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
     313           0 :                                 free(map->map_256tb.map[i]);
     314           0 :                         }
     315           0 :                         free(map);
     316           0 :                         return NULL;
     317             :                 }
     318           0 :                 TAILQ_INSERT_TAIL(&g_spdk_mem_maps, map, tailq);
     319           0 :                 pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     320           0 :         }
     321             : 
     322           0 :         return map;
     323           0 : }
     324             : 
     325             : void
     326           0 : spdk_mem_map_free(struct spdk_mem_map **pmap)
     327             : {
     328             :         struct spdk_mem_map *map;
     329             :         size_t i;
     330             : 
     331           0 :         if (!pmap) {
     332           0 :                 return;
     333             :         }
     334             : 
     335           0 :         map = *pmap;
     336             : 
     337           0 :         if (!map) {
     338           0 :                 return;
     339             :         }
     340             : 
     341           0 :         if (map->ops.notify_cb) {
     342           0 :                 pthread_mutex_lock(&g_spdk_mem_map_mutex);
     343           0 :                 mem_map_notify_walk(map, SPDK_MEM_MAP_NOTIFY_UNREGISTER);
     344           0 :                 TAILQ_REMOVE(&g_spdk_mem_maps, map, tailq);
     345           0 :                 pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     346           0 :         }
     347             : 
     348           0 :         for (i = 0; i < sizeof(map->map_256tb.map) / sizeof(map->map_256tb.map[0]); i++) {
     349           0 :                 free(map->map_256tb.map[i]);
     350           0 :         }
     351             : 
     352           0 :         pthread_mutex_destroy(&map->mutex);
     353             : 
     354           0 :         free(map);
     355           0 :         *pmap = NULL;
     356           0 : }
     357             : 
     358             : uint64_t
     359           0 : spdk_mem_map_get_page_size(void)
     360             : {
     361           0 :         return g_map_page_cfg.size;
     362             : }
     363             : 
     364             : int
     365           0 : spdk_mem_register(void *_vaddr, size_t len)
     366             : {
     367             :         struct spdk_mem_map *map;
     368             :         int rc;
     369           0 :         uint64_t vaddr = (uintptr_t)_vaddr;
     370             :         uint64_t seg_vaddr;
     371             :         size_t seg_len;
     372             :         uint64_t reg;
     373             : 
     374           0 :         if ((uintptr_t)vaddr & ~MASK_256TB) {
     375           0 :                 DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
     376           0 :                 return -EINVAL;
     377             :         }
     378             : 
     379           0 :         if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
     380           0 :                 DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
     381             :                             __func__, vaddr, len);
     382           0 :                 return -EINVAL;
     383             :         }
     384             : 
     385           0 :         if (len == 0) {
     386           0 :                 return 0;
     387             :         }
     388             : 
     389           0 :         pthread_mutex_lock(&g_spdk_mem_map_mutex);
     390             : 
     391           0 :         seg_vaddr = vaddr;
     392           0 :         seg_len = len;
     393           0 :         while (seg_len > 0) {
     394           0 :                 reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
     395           0 :                 if (reg & REG_MAP_REGISTERED) {
     396           0 :                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     397           0 :                         return -EBUSY;
     398             :                 }
     399           0 :                 seg_vaddr += MAP_PAGE_SIZE;
     400           0 :                 seg_len -= MAP_PAGE_SIZE;
     401             :         }
     402             : 
     403           0 :         seg_vaddr = vaddr;
     404           0 :         seg_len = 0;
     405           0 :         while (len > 0) {
     406           0 :                 spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, MAP_PAGE_SIZE,
     407           0 :                                              seg_len == 0 ? REG_MAP_REGISTERED | REG_MAP_NOTIFY_START : REG_MAP_REGISTERED);
     408           0 :                 seg_len += MAP_PAGE_SIZE;
     409           0 :                 vaddr += MAP_PAGE_SIZE;
     410           0 :                 len -= MAP_PAGE_SIZE;
     411             :         }
     412             : 
     413           0 :         TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
     414           0 :                 rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_REGISTER,
     415           0 :                                         (void *)seg_vaddr, seg_len);
     416           0 :                 if (rc != 0) {
     417           0 :                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     418           0 :                         return rc;
     419             :                 }
     420           0 :         }
     421             : 
     422           0 :         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     423           0 :         return 0;
     424           0 : }
     425             : 
     426             : int
     427           0 : spdk_mem_unregister(void *_vaddr, size_t len)
     428             : {
     429             :         struct spdk_mem_map *map;
     430             :         int rc;
     431           0 :         uint64_t vaddr = (uintptr_t)_vaddr;
     432             :         uint64_t seg_vaddr;
     433             :         size_t seg_len;
     434             :         uint64_t reg, newreg;
     435             : 
     436           0 :         if ((uintptr_t)vaddr & ~MASK_256TB) {
     437           0 :                 DEBUG_PRINT("invalid usermode virtual address %jx\n", vaddr);
     438           0 :                 return -EINVAL;
     439             :         }
     440             : 
     441           0 :         if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
     442           0 :                 DEBUG_PRINT("invalid %s parameters, vaddr=%jx len=%ju\n",
     443             :                             __func__, vaddr, len);
     444           0 :                 return -EINVAL;
     445             :         }
     446             : 
     447           0 :         pthread_mutex_lock(&g_spdk_mem_map_mutex);
     448             : 
     449             :         /* The first page must be a start of a region. Also check if it's
     450             :          * registered to make sure we don't return -ERANGE for non-registered
     451             :          * regions.
     452             :          */
     453           0 :         reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
     454           0 :         if ((reg & REG_MAP_REGISTERED) && (reg & REG_MAP_NOTIFY_START) == 0) {
     455           0 :                 pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     456           0 :                 return -ERANGE;
     457             :         }
     458             : 
     459           0 :         seg_vaddr = vaddr;
     460           0 :         seg_len = len;
     461           0 :         while (seg_len > 0) {
     462           0 :                 reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
     463           0 :                 if ((reg & REG_MAP_REGISTERED) == 0) {
     464           0 :                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     465           0 :                         return -EINVAL;
     466             :                 }
     467           0 :                 seg_vaddr += MAP_PAGE_SIZE;
     468           0 :                 seg_len -= MAP_PAGE_SIZE;
     469             :         }
     470             : 
     471           0 :         newreg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
     472             :         /* If the next page is registered, it must be a start of a region as well,
     473             :          * otherwise we'd be unregistering only a part of a region.
     474             :          */
     475           0 :         if ((newreg & REG_MAP_NOTIFY_START) == 0 && (newreg & REG_MAP_REGISTERED)) {
     476           0 :                 pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     477           0 :                 return -ERANGE;
     478             :         }
     479           0 :         seg_vaddr = vaddr;
     480           0 :         seg_len = 0;
     481             : 
     482           0 :         while (len > 0) {
     483           0 :                 reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)vaddr, NULL);
     484           0 :                 spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, MAP_PAGE_SIZE, 0);
     485             : 
     486           0 :                 if (seg_len > 0 && (reg & REG_MAP_NOTIFY_START)) {
     487           0 :                         TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
     488           0 :                                 rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
     489           0 :                                                         (void *)seg_vaddr, seg_len);
     490           0 :                                 if (rc != 0) {
     491           0 :                                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     492           0 :                                         return rc;
     493             :                                 }
     494           0 :                         }
     495             : 
     496           0 :                         seg_vaddr = vaddr;
     497           0 :                         seg_len = MAP_PAGE_SIZE;
     498           0 :                 } else {
     499           0 :                         seg_len += MAP_PAGE_SIZE;
     500             :                 }
     501             : 
     502           0 :                 vaddr += MAP_PAGE_SIZE;
     503           0 :                 len -= MAP_PAGE_SIZE;
     504             :         }
     505             : 
     506           0 :         if (seg_len > 0) {
     507           0 :                 TAILQ_FOREACH_REVERSE(map, &g_spdk_mem_maps, spdk_mem_map_head, tailq) {
     508           0 :                         rc = map->ops.notify_cb(map->cb_ctx, map, SPDK_MEM_MAP_NOTIFY_UNREGISTER,
     509           0 :                                                 (void *)seg_vaddr, seg_len);
     510           0 :                         if (rc != 0) {
     511           0 :                                 pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     512           0 :                                 return rc;
     513             :                         }
     514           0 :                 }
     515           0 :         }
     516             : 
     517           0 :         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     518           0 :         return 0;
     519           0 : }
     520             : 
     521             : int
     522           0 : spdk_mem_reserve(void *vaddr, size_t len)
     523             : {
     524             :         struct spdk_mem_map *map;
     525             :         void *seg_vaddr;
     526             :         size_t seg_len;
     527             :         uint64_t reg;
     528             : 
     529           0 :         if ((uintptr_t)vaddr & ~MASK_256TB) {
     530           0 :                 DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
     531           0 :                 return -EINVAL;
     532             :         }
     533             : 
     534           0 :         if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (len & MAP_PAGE_MASK)) {
     535           0 :                 DEBUG_PRINT("invalid %s parameters, vaddr=%p len=%ju\n",
     536             :                             __func__, vaddr, len);
     537           0 :                 return -EINVAL;
     538             :         }
     539             : 
     540           0 :         if (len == 0) {
     541           0 :                 return 0;
     542             :         }
     543             : 
     544           0 :         pthread_mutex_lock(&g_spdk_mem_map_mutex);
     545             : 
     546             :         /* Check if any part of this range is already registered */
     547           0 :         seg_vaddr = vaddr;
     548           0 :         seg_len = len;
     549           0 :         while (seg_len > 0) {
     550           0 :                 reg = spdk_mem_map_translate(g_mem_reg_map, (uint64_t)seg_vaddr, NULL);
     551           0 :                 if (reg & REG_MAP_REGISTERED) {
     552           0 :                         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     553           0 :                         return -EBUSY;
     554             :                 }
     555           0 :                 seg_vaddr += MAP_PAGE_SIZE;
     556           0 :                 seg_len -= MAP_PAGE_SIZE;
     557             :         }
     558             : 
     559             :         /* Simply set the translation to the memory map's default. This allocates the space in the
     560             :          * map but does not provide a valid translation. */
     561           0 :         spdk_mem_map_set_translation(g_mem_reg_map, (uint64_t)vaddr, len,
     562           0 :                                      g_mem_reg_map->default_translation);
     563             : 
     564           0 :         TAILQ_FOREACH(map, &g_spdk_mem_maps, tailq) {
     565           0 :                 spdk_mem_map_set_translation(map, (uint64_t)vaddr, len, map->default_translation);
     566           0 :         }
     567             : 
     568           0 :         pthread_mutex_unlock(&g_spdk_mem_map_mutex);
     569           0 :         return 0;
     570           0 : }
     571             : 
     572             : static struct map_1gb *
     573           0 : mem_map_get_map_1gb(struct spdk_mem_map *map, uint64_t vfn_page)
     574             : {
     575             :         struct map_1gb *map_1gb;
     576           0 :         uint64_t idx_256tb = MAP_256TB_IDX(vfn_page);
     577             :         size_t i;
     578             : 
     579           0 :         if (spdk_unlikely(idx_256tb >= SPDK_COUNTOF(map->map_256tb.map))) {
     580           0 :                 return NULL;
     581             :         }
     582             : 
     583           0 :         map_1gb = map->map_256tb.map[idx_256tb];
     584             : 
     585           0 :         if (!map_1gb) {
     586           0 :                 pthread_mutex_lock(&map->mutex);
     587             : 
     588             :                 /* Recheck to make sure nobody else got the mutex first. */
     589           0 :                 map_1gb = map->map_256tb.map[idx_256tb];
     590           0 :                 if (!map_1gb) {
     591           0 :                         map_1gb = malloc(MAP_SIZE_OF_MAP_1GB);
     592           0 :                         if (map_1gb) {
     593             :                                 /* initialize all entries to default translation */
     594           0 :                                 for (i = 0; i < MAP_NUM_PAGES_PER_GB; i++) {
     595           0 :                                         map_1gb->map[i].translation = map->default_translation;
     596           0 :                                 }
     597           0 :                                 map->map_256tb.map[idx_256tb] = map_1gb;
     598           0 :                         }
     599           0 :                 }
     600             : 
     601           0 :                 pthread_mutex_unlock(&map->mutex);
     602             : 
     603           0 :                 if (!map_1gb) {
     604           0 :                         DEBUG_PRINT("allocation failed\n");
     605           0 :                         return NULL;
     606             :                 }
     607           0 :         }
     608             : 
     609           0 :         return map_1gb;
     610           0 : }
     611             : 
     612             : int
     613           0 : spdk_mem_map_set_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size,
     614             :                              uint64_t translation)
     615             : {
     616             :         uint64_t vfn_page;
     617             :         struct map_1gb *map_1gb;
     618             :         uint64_t idx_1gb;
     619             :         struct map_page *map_page;
     620             : 
     621           0 :         if ((uintptr_t)vaddr & ~MASK_256TB) {
     622           0 :                 DEBUG_PRINT("invalid usermode virtual address %" PRIu64 "\n", vaddr);
     623           0 :                 return -EINVAL;
     624             :         }
     625             : 
     626             :         /* Only page-aligned registrations are supported */
     627           0 :         if (((uintptr_t)vaddr & MAP_PAGE_MASK) || (size & MAP_PAGE_MASK)) {
     628           0 :                 DEBUG_PRINT("invalid %s parameters, vaddr=%" PRIu64 " len=%" PRIu64 "\n",
     629             :                             __func__, vaddr, size);
     630           0 :                 return -EINVAL;
     631             :         }
     632             : 
     633           0 :         vfn_page = vaddr >> MAP_PAGE_SHIFT;
     634             : 
     635           0 :         while (size) {
     636           0 :                 map_1gb = mem_map_get_map_1gb(map, vfn_page);
     637           0 :                 if (!map_1gb) {
     638           0 :                         DEBUG_PRINT("could not get %p map\n", (void *)vaddr);
     639           0 :                         return -ENOMEM;
     640             :                 }
     641             : 
     642           0 :                 idx_1gb = MAP_1GB_IDX(vfn_page);
     643           0 :                 map_page = &map_1gb->map[idx_1gb];
     644           0 :                 map_page->translation = translation;
     645             : 
     646           0 :                 size -= MAP_PAGE_SIZE;
     647           0 :                 vfn_page++;
     648             :         }
     649             : 
     650           0 :         return 0;
     651           0 : }
     652             : 
     653             : int
     654           0 : spdk_mem_map_clear_translation(struct spdk_mem_map *map, uint64_t vaddr, uint64_t size)
     655             : {
     656           0 :         return spdk_mem_map_set_translation(map, vaddr, size, map->default_translation);
     657             : }
     658             : 
     659             : inline uint64_t
     660           0 : spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
     661             : {
     662             :         const struct map_1gb *map_1gb;
     663             :         const struct map_page *map_page;
     664             :         uint64_t idx_256tb;
     665             :         uint64_t idx_1gb;
     666             :         uint64_t vfn_page;
     667             :         uint64_t cur_size;
     668             :         uint64_t prev_translation;
     669             :         uint64_t orig_translation;
     670             : 
     671           0 :         if (spdk_unlikely(vaddr & ~MASK_256TB)) {
     672           0 :                 DEBUG_PRINT("invalid usermode virtual address %p\n", (void *)vaddr);
     673           0 :                 return map->default_translation;
     674             :         }
     675             : 
     676           0 :         vfn_page = vaddr >> MAP_PAGE_SHIFT;
     677           0 :         idx_256tb = MAP_256TB_IDX(vfn_page);
     678           0 :         idx_1gb = MAP_1GB_IDX(vfn_page);
     679             : 
     680           0 :         map_1gb = map->map_256tb.map[idx_256tb];
     681           0 :         if (spdk_unlikely(!map_1gb)) {
     682           0 :                 return map->default_translation;
     683             :         }
     684             : 
     685           0 :         cur_size = MAP_PAGE_SIZE - MAP_PAGE_OFFSET(vaddr);
     686           0 :         map_page = &map_1gb->map[idx_1gb];
     687           0 :         if (size == NULL || map->ops.are_contiguous == NULL ||
     688           0 :             map_page->translation == map->default_translation) {
     689           0 :                 if (size != NULL) {
     690           0 :                         *size = spdk_min(*size, cur_size);
     691           0 :                 }
     692           0 :                 return map_page->translation;
     693             :         }
     694             : 
     695           0 :         orig_translation = map_page->translation;
     696           0 :         prev_translation = orig_translation;
     697           0 :         while (cur_size < *size) {
     698           0 :                 vfn_page++;
     699           0 :                 idx_256tb = MAP_256TB_IDX(vfn_page);
     700           0 :                 idx_1gb = MAP_1GB_IDX(vfn_page);
     701             : 
     702           0 :                 map_1gb = map->map_256tb.map[idx_256tb];
     703           0 :                 if (spdk_unlikely(!map_1gb)) {
     704           0 :                         break;
     705             :                 }
     706             : 
     707           0 :                 map_page = &map_1gb->map[idx_1gb];
     708           0 :                 if (!map->ops.are_contiguous(prev_translation, map_page->translation)) {
     709           0 :                         break;
     710             :                 }
     711             : 
     712           0 :                 cur_size += MAP_PAGE_SIZE;
     713           0 :                 prev_translation = map_page->translation;
     714             :         }
     715             : 
     716           0 :         *size = spdk_min(*size, cur_size);
     717           0 :         return orig_translation;
     718           0 : }
     719             : 
     720             : static void
     721           0 : memory_hotplug_cb(enum rte_mem_event event_type,
     722             :                   const void *addr, size_t len, void *arg)
     723             : {
     724           0 :         if (event_type == RTE_MEM_EVENT_ALLOC) {
     725           0 :                 spdk_mem_register((void *)addr, len);
     726             : 
     727           0 :                 if (!spdk_env_dpdk_external_init()) {
     728           0 :                         return;
     729             :                 }
     730             : 
     731             :                 /* When the user initialized DPDK separately, we can't
     732             :                  * be sure that --match-allocations RTE flag was specified.
     733             :                  * Without this flag, DPDK can free memory in different units
     734             :                  * than it was allocated. It doesn't work with things like RDMA MRs.
     735             :                  *
     736             :                  * For such cases, we mark segments so they aren't freed.
     737             :                  */
     738           0 :                 while (len > 0) {
     739             :                         struct rte_memseg *seg;
     740             : 
     741           0 :                         seg = rte_mem_virt2memseg(addr, NULL);
     742           0 :                         assert(seg != NULL);
     743           0 :                         seg->flags |= RTE_MEMSEG_FLAG_DO_NOT_FREE;
     744           0 :                         addr = (void *)((uintptr_t)addr + seg->hugepage_sz);
     745           0 :                         len -= seg->hugepage_sz;
     746             :                 }
     747           0 :         } else if (event_type == RTE_MEM_EVENT_FREE) {
     748           0 :                 spdk_mem_unregister((void *)addr, len);
     749           0 :         }
     750           0 : }
     751             : 
     752             : static int
     753           0 : memory_iter_cb(const struct rte_memseg_list *msl,
     754             :                const struct rte_memseg *ms, size_t len, void *arg)
     755             : {
     756           0 :         return spdk_mem_register(ms->addr, len);
     757             : }
     758             : 
     759             : static bool g_mem_event_cb_registered = false;
     760             : 
     761             : static int
     762           0 : mem_map_mem_event_callback_register(void)
     763             : {
     764             :         int rc;
     765             : 
     766           0 :         rc = rte_mem_event_callback_register("spdk", memory_hotplug_cb, NULL);
     767           0 :         if (rc != 0) {
     768           0 :                 return rc;
     769             :         }
     770             : 
     771           0 :         g_mem_event_cb_registered = true;
     772           0 :         return 0;
     773           0 : }
     774             : 
     775             : static void
     776           0 : mem_map_mem_event_callback_unregister(void)
     777             : {
     778           0 :         if (g_mem_event_cb_registered) {
     779           0 :                 g_mem_event_cb_registered = false;
     780           0 :                 rte_mem_event_callback_unregister("spdk", NULL);
     781           0 :         }
     782           0 : }
     783             : 
     784             : int
     785           0 : mem_map_init(bool legacy_mem)
     786             : {
     787             :         int rc;
     788             : 
     789           0 :         g_legacy_mem = legacy_mem;
     790             : 
     791           0 :         g_mem_reg_map = spdk_mem_map_alloc(0, NULL, NULL);
     792           0 :         if (g_mem_reg_map == NULL) {
     793           0 :                 DEBUG_PRINT("memory registration map allocation failed\n");
     794           0 :                 return -ENOMEM;
     795             :         }
     796             : 
     797           0 :         if (!g_legacy_mem) {
     798             :                 /**
     799             :                  * To prevent DPDK complaining, only register the callback when
     800             :                  * we are not in legacy mem mode.
     801             :                  */
     802           0 :                 rc = mem_map_mem_event_callback_register();
     803           0 :                 if (rc != 0) {
     804           0 :                         DEBUG_PRINT("memory event callback registration failed, rc = %d\n", rc);
     805           0 :                         goto err_free_reg_map;
     806             :                 }
     807           0 :         }
     808             : 
     809             :         /*
     810             :          * Walk all DPDK memory segments and register them
     811             :          * with the main memory map
     812             :          */
     813           0 :         rc = rte_memseg_contig_walk(memory_iter_cb, NULL);
     814           0 :         if (rc != 0) {
     815           0 :                 DEBUG_PRINT("memory segments walking failed, rc = %d\n", rc);
     816           0 :                 goto err_unregister_mem_cb;
     817             :         }
     818             : 
     819           0 :         return 0;
     820             : 
     821             : err_unregister_mem_cb:
     822           0 :         mem_map_mem_event_callback_unregister();
     823             : err_free_reg_map:
     824           0 :         spdk_mem_map_free(&g_mem_reg_map);
     825           0 :         return rc;
     826           0 : }
     827             : 
     828             : void
     829           0 : mem_map_fini(void)
     830             : {
     831           0 :         mem_map_mem_event_callback_unregister();
     832           0 :         spdk_mem_map_free(&g_mem_reg_map);
     833           0 : }
     834             : 
     835             : bool
     836           0 : spdk_iommu_is_enabled(void)
     837             : {
     838             : #if VFIO_ENABLED
     839             :         return g_vfio.enabled && !g_vfio.noiommu_enabled;
     840             : #else
     841           0 :         return false;
     842             : #endif
     843             : }
     844             : 
     845             : struct spdk_vtophys_pci_device {
     846             :         struct rte_pci_device *pci_device;
     847             :         TAILQ_ENTRY(spdk_vtophys_pci_device) tailq;
     848             : };
     849             : 
     850             : static pthread_mutex_t g_vtophys_pci_devices_mutex = PTHREAD_MUTEX_INITIALIZER;
     851             : static TAILQ_HEAD(, spdk_vtophys_pci_device) g_vtophys_pci_devices =
     852             :         TAILQ_HEAD_INITIALIZER(g_vtophys_pci_devices);
     853             : 
     854             : static struct spdk_mem_map *g_vtophys_map;
     855             : static struct spdk_mem_map *g_phys_ref_map;
     856             : static struct spdk_mem_map *g_numa_map;
     857             : 
     858             : #if VFIO_ENABLED
     859             : static int
     860             : _vfio_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
     861             : {
     862             :         struct spdk_vfio_dma_map *dma_map;
     863             :         int ret;
     864             : 
     865             :         dma_map = calloc(1, sizeof(*dma_map));
     866             :         if (dma_map == NULL) {
     867             :                 return -ENOMEM;
     868             :         }
     869             : 
     870             :         dma_map->map.argsz = sizeof(dma_map->map);
     871             :         dma_map->map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
     872             :         dma_map->map.vaddr = vaddr;
     873             :         dma_map->map.iova = iova;
     874             :         dma_map->map.size = size;
     875             : 
     876             :         if (g_vfio.device_ref == 0) {
     877             :                 /* VFIO requires at least one device (IOMMU group) to be added to
     878             :                  * a VFIO container before it is possible to perform any IOMMU
     879             :                  * operations on that container. This memory will be mapped once
     880             :                  * the first device (IOMMU group) is hotplugged.
     881             :                  *
     882             :                  * Since the vfio container is managed internally by DPDK, it is
     883             :                  * also possible that some device is already in that container, but
     884             :                  * it's not managed by SPDK -  e.g. an NIC attached internally
     885             :                  * inside DPDK. We could map the memory straight away in such
     886             :                  * scenario, but there's no need to do it. DPDK devices clearly
     887             :                  * don't need our mappings and hence we defer the mapping
     888             :                  * unconditionally until the first SPDK-managed device is
     889             :                  * hotplugged.
     890             :                  */
     891             :                 goto out_insert;
     892             :         }
     893             : 
     894             :         ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
     895             :         if (ret) {
     896             :                 /* There are cases the vfio container doesn't have IOMMU group, it's safe for this case */
     897             :                 SPDK_NOTICELOG("Cannot set up DMA mapping, error %d, ignored\n", errno);
     898             :         }
     899             : 
     900             : out_insert:
     901             :         TAILQ_INSERT_TAIL(&g_vfio.maps, dma_map, tailq);
     902             :         return 0;
     903             : }
     904             : 
     905             : 
     906             : static int
     907             : vtophys_iommu_map_dma(uint64_t vaddr, uint64_t iova, uint64_t size)
     908             : {
     909             :         uint64_t refcount;
     910             :         int ret;
     911             : 
     912             :         refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
     913             :         assert(refcount < UINT64_MAX);
     914             :         if (refcount > 0) {
     915             :                 spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
     916             :                 return 0;
     917             :         }
     918             : 
     919             :         pthread_mutex_lock(&g_vfio.mutex);
     920             :         ret = _vfio_iommu_map_dma(vaddr, iova, size);
     921             :         pthread_mutex_unlock(&g_vfio.mutex);
     922             :         if (ret) {
     923             :                 return ret;
     924             :         }
     925             : 
     926             :         spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount + 1);
     927             :         return 0;
     928             : }
     929             : 
     930             : int
     931             : vtophys_iommu_map_dma_bar(uint64_t vaddr, uint64_t iova, uint64_t size)
     932             : {
     933             :         int ret;
     934             : 
     935             :         pthread_mutex_lock(&g_vfio.mutex);
     936             :         ret = _vfio_iommu_map_dma(vaddr, iova, size);
     937             :         pthread_mutex_unlock(&g_vfio.mutex);
     938             : 
     939             :         return ret;
     940             : }
     941             : 
     942             : static int
     943             : _vfio_iommu_unmap_dma(struct spdk_vfio_dma_map *dma_map)
     944             : {
     945             :         struct vfio_iommu_type1_dma_unmap unmap = {};
     946             :         int ret;
     947             : 
     948             :         if (g_vfio.device_ref == 0) {
     949             :                 /* Memory is not mapped anymore, just remove it's references */
     950             :                 goto out_remove;
     951             :         }
     952             : 
     953             :         unmap.argsz = sizeof(unmap);
     954             :         unmap.flags = 0;
     955             :         unmap.iova = dma_map->map.iova;
     956             :         unmap.size = dma_map->map.size;
     957             :         ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
     958             :         if (ret) {
     959             :                 SPDK_NOTICELOG("Cannot clear DMA mapping, error %d, ignored\n", errno);
     960             :         }
     961             : 
     962             : out_remove:
     963             :         TAILQ_REMOVE(&g_vfio.maps, dma_map, tailq);
     964             :         free(dma_map);
     965             :         return 0;
     966             : }
     967             : 
     968             : static int
     969             : vtophys_iommu_unmap_dma(uint64_t iova, uint64_t size)
     970             : {
     971             :         struct spdk_vfio_dma_map *dma_map;
     972             :         uint64_t refcount;
     973             :         int ret;
     974             : 
     975             :         pthread_mutex_lock(&g_vfio.mutex);
     976             :         TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
     977             :                 if (dma_map->map.iova == iova) {
     978             :                         break;
     979             :                 }
     980             :         }
     981             : 
     982             :         if (dma_map == NULL) {
     983             :                 DEBUG_PRINT("Cannot clear DMA mapping for IOVA %"PRIx64" - it's not mapped\n", iova);
     984             :                 pthread_mutex_unlock(&g_vfio.mutex);
     985             :                 return -ENXIO;
     986             :         }
     987             : 
     988             :         refcount = spdk_mem_map_translate(g_phys_ref_map, iova, NULL);
     989             :         assert(refcount < UINT64_MAX);
     990             :         if (refcount > 0) {
     991             :                 spdk_mem_map_set_translation(g_phys_ref_map, iova, size, refcount - 1);
     992             :         }
     993             : 
     994             :         /* We still have outstanding references, don't clear it. */
     995             :         if (refcount > 1) {
     996             :                 pthread_mutex_unlock(&g_vfio.mutex);
     997             :                 return 0;
     998             :         }
     999             : 
    1000             :         /** don't support partial or multiple-page unmap for now */
    1001             :         assert(dma_map->map.size == size);
    1002             : 
    1003             :         ret = _vfio_iommu_unmap_dma(dma_map);
    1004             :         pthread_mutex_unlock(&g_vfio.mutex);
    1005             : 
    1006             :         return ret;
    1007             : }
    1008             : 
    1009             : int
    1010             : vtophys_iommu_unmap_dma_bar(uint64_t vaddr)
    1011             : {
    1012             :         struct spdk_vfio_dma_map *dma_map;
    1013             :         int ret;
    1014             : 
    1015             :         pthread_mutex_lock(&g_vfio.mutex);
    1016             :         TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
    1017             :                 if (dma_map->map.vaddr == vaddr) {
    1018             :                         break;
    1019             :                 }
    1020             :         }
    1021             : 
    1022             :         if (dma_map == NULL) {
    1023             :                 DEBUG_PRINT("Cannot clear DMA mapping for address %"PRIx64" - it's not mapped\n", vaddr);
    1024             :                 pthread_mutex_unlock(&g_vfio.mutex);
    1025             :                 return -ENXIO;
    1026             :         }
    1027             : 
    1028             :         ret = _vfio_iommu_unmap_dma(dma_map);
    1029             :         pthread_mutex_unlock(&g_vfio.mutex);
    1030             :         return ret;
    1031             : }
    1032             : #endif
    1033             : 
    1034             : static uint64_t
    1035           0 : vtophys_get_paddr_memseg(uint64_t vaddr)
    1036             : {
    1037             :         uintptr_t paddr;
    1038             :         struct rte_memseg *seg;
    1039             : 
    1040           0 :         seg = rte_mem_virt2memseg((void *)(uintptr_t)vaddr, NULL);
    1041           0 :         if (seg != NULL) {
    1042           0 :                 paddr = seg->iova;
    1043           0 :                 if (paddr == RTE_BAD_IOVA) {
    1044           0 :                         return SPDK_VTOPHYS_ERROR;
    1045             :                 }
    1046           0 :                 paddr += (vaddr - (uintptr_t)seg->addr);
    1047           0 :                 return paddr;
    1048             :         }
    1049             : 
    1050           0 :         return SPDK_VTOPHYS_ERROR;
    1051           0 : }
    1052             : 
    1053             : /* Try to get the paddr from /proc/self/pagemap */
    1054             : static uint64_t
    1055           0 : vtophys_get_paddr_pagemap(uint64_t vaddr)
    1056             : {
    1057             :         uintptr_t paddr;
    1058             : 
    1059             :         /* Silence static analyzers */
    1060           0 :         assert(vaddr != 0);
    1061           0 :         paddr = rte_mem_virt2iova((void *)vaddr);
    1062           0 :         if (paddr == RTE_BAD_IOVA) {
    1063             :                 /*
    1064             :                  * The vaddr may be valid but doesn't have a backing page
    1065             :                  * assigned yet.  Touch the page to ensure a backing page
    1066             :                  * gets assigned, then try to translate again.
    1067             :                  */
    1068           0 :                 rte_atomic64_read((rte_atomic64_t *)vaddr);
    1069           0 :                 paddr = rte_mem_virt2iova((void *)vaddr);
    1070           0 :         }
    1071           0 :         if (paddr == RTE_BAD_IOVA) {
    1072             :                 /* Unable to get to the physical address. */
    1073           0 :                 return SPDK_VTOPHYS_ERROR;
    1074             :         }
    1075             : 
    1076           0 :         return paddr;
    1077           0 : }
    1078             : 
    1079             : static uint64_t
    1080           0 : pci_device_vtophys(struct rte_pci_device *dev, uint64_t vaddr, size_t len)
    1081             : {
    1082             :         struct rte_mem_resource *res;
    1083             :         uint64_t paddr;
    1084             :         unsigned r;
    1085             : 
    1086           0 :         for (r = 0; r < PCI_MAX_RESOURCE; r++) {
    1087           0 :                 res = dpdk_pci_device_get_mem_resource(dev, r);
    1088             : 
    1089           0 :                 if (res->phys_addr == 0 || vaddr < (uint64_t)res->addr ||
    1090           0 :                     (vaddr + len) >= (uint64_t)res->addr + res->len) {
    1091           0 :                         continue;
    1092             :                 }
    1093             : 
    1094             : #if VFIO_ENABLED
    1095             :                 if (spdk_iommu_is_enabled() && rte_eal_iova_mode() == RTE_IOVA_VA) {
    1096             :                         /*
    1097             :                          * The IOMMU is on and we're using IOVA == VA. The BAR was
    1098             :                          * automatically registered when it was mapped, so just return
    1099             :                          * the virtual address here.
    1100             :                          */
    1101             :                         return vaddr;
    1102             :                 }
    1103             : #endif
    1104           0 :                 paddr = res->phys_addr + (vaddr - (uint64_t)res->addr);
    1105           0 :                 return paddr;
    1106             :         }
    1107             : 
    1108           0 :         return SPDK_VTOPHYS_ERROR;
    1109           0 : }
    1110             : 
    1111             : /* Try to get the paddr from pci devices */
    1112             : static uint64_t
    1113           0 : vtophys_get_paddr_pci(uint64_t vaddr, size_t len)
    1114             : {
    1115             :         struct spdk_vtophys_pci_device *vtophys_dev;
    1116             :         uintptr_t paddr;
    1117             :         struct rte_pci_device   *dev;
    1118             : 
    1119           0 :         pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
    1120           0 :         TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
    1121           0 :                 dev = vtophys_dev->pci_device;
    1122           0 :                 paddr = pci_device_vtophys(dev, vaddr, len);
    1123           0 :                 if (paddr != SPDK_VTOPHYS_ERROR) {
    1124           0 :                         pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
    1125           0 :                         return paddr;
    1126             :                 }
    1127           0 :         }
    1128           0 :         pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
    1129             : 
    1130           0 :         return SPDK_VTOPHYS_ERROR;
    1131           0 : }
    1132             : 
    1133             : static int
    1134           0 : vtophys_notify(void *cb_ctx, struct spdk_mem_map *map,
    1135             :                enum spdk_mem_map_notify_action action,
    1136             :                void *vaddr, size_t len)
    1137             : {
    1138           0 :         int rc = 0;
    1139             :         uint64_t paddr;
    1140             : 
    1141           0 :         if ((uintptr_t)vaddr & ~MASK_256TB) {
    1142           0 :                 DEBUG_PRINT("invalid usermode virtual address %p\n", vaddr);
    1143           0 :                 return -EINVAL;
    1144             :         }
    1145             : 
    1146           0 :         if (((uintptr_t)vaddr & MASK_2MB) || (len & MASK_2MB)) {
    1147           0 :                 DEBUG_PRINT("invalid parameters, vaddr=%p len=%ju\n",
    1148             :                             vaddr, len);
    1149           0 :                 return -EINVAL;
    1150             :         }
    1151             : 
    1152             :         /* Get the physical address from the DPDK memsegs */
    1153           0 :         paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
    1154             : 
    1155           0 :         switch (action) {
    1156             :         case SPDK_MEM_MAP_NOTIFY_REGISTER:
    1157           0 :                 if (paddr == SPDK_VTOPHYS_ERROR) {
    1158             :                         /* This is not an address that DPDK is managing. */
    1159             : 
    1160             :                         /* Check if this is a PCI BAR. They need special handling */
    1161           0 :                         paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
    1162           0 :                         if (paddr != SPDK_VTOPHYS_ERROR) {
    1163             :                                 /* Get paddr for each 2MB chunk in this address range */
    1164           0 :                                 while (len > 0) {
    1165           0 :                                         paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
    1166           0 :                                         if (paddr == SPDK_VTOPHYS_ERROR) {
    1167           0 :                                                 DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
    1168           0 :                                                 return -EFAULT;
    1169             :                                         }
    1170             : 
    1171           0 :                                         rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
    1172           0 :                                         if (rc != 0) {
    1173           0 :                                                 return rc;
    1174             :                                         }
    1175             : 
    1176           0 :                                         vaddr += VALUE_2MB;
    1177           0 :                                         len -= VALUE_2MB;
    1178             :                                 }
    1179             : 
    1180           0 :                                 return 0;
    1181             :                         }
    1182             : 
    1183             : #if VFIO_ENABLED
    1184             :                         enum rte_iova_mode iova_mode;
    1185             : 
    1186             :                         iova_mode = rte_eal_iova_mode();
    1187             : 
    1188             :                         if (spdk_iommu_is_enabled() && iova_mode == RTE_IOVA_VA) {
    1189             :                                 /* We'll use the virtual address as the iova to match DPDK. */
    1190             :                                 paddr = (uint64_t)vaddr;
    1191             :                                 rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, len);
    1192             :                                 if (rc) {
    1193             :                                         return -EFAULT;
    1194             :                                 }
    1195             :                                 while (len > 0) {
    1196             :                                         rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
    1197             :                                         if (rc != 0) {
    1198             :                                                 return rc;
    1199             :                                         }
    1200             :                                         vaddr += VALUE_2MB;
    1201             :                                         paddr += VALUE_2MB;
    1202             :                                         len -= VALUE_2MB;
    1203             :                                 }
    1204             :                         } else
    1205             : #endif
    1206             :                         {
    1207             :                                 /* Get the physical address from /proc/self/pagemap. */
    1208           0 :                                 paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
    1209           0 :                                 if (paddr == SPDK_VTOPHYS_ERROR) {
    1210           0 :                                         DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
    1211           0 :                                         return -EFAULT;
    1212             :                                 }
    1213             : 
    1214             :                                 /* Get paddr for each 2MB chunk in this address range */
    1215           0 :                                 while (len > 0) {
    1216             :                                         /* Get the physical address from /proc/self/pagemap. */
    1217           0 :                                         paddr = vtophys_get_paddr_pagemap((uint64_t)vaddr);
    1218             : 
    1219           0 :                                         if (paddr == SPDK_VTOPHYS_ERROR) {
    1220           0 :                                                 DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
    1221           0 :                                                 return -EFAULT;
    1222             :                                         }
    1223             : 
    1224           0 :                                         if (paddr & MASK_2MB) {
    1225           0 :                                                 DEBUG_PRINT("invalid paddr 0x%" PRIx64 " - must be 2MB aligned\n", paddr);
    1226           0 :                                                 return -EINVAL;
    1227             :                                         }
    1228             : #if VFIO_ENABLED
    1229             :                                         /* If the IOMMU is on, but DPDK is using iova-mode=pa, we want to register this memory
    1230             :                                          * with the IOMMU using the physical address to match. */
    1231             :                                         if (spdk_iommu_is_enabled()) {
    1232             :                                                 rc = vtophys_iommu_map_dma((uint64_t)vaddr, paddr, VALUE_2MB);
    1233             :                                                 if (rc) {
    1234             :                                                         DEBUG_PRINT("Unable to assign vaddr %p to paddr 0x%" PRIx64 "\n", vaddr, paddr);
    1235             :                                                         return -EFAULT;
    1236             :                                                 }
    1237             :                                         }
    1238             : #endif
    1239             : 
    1240           0 :                                         rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
    1241           0 :                                         if (rc != 0) {
    1242           0 :                                                 return rc;
    1243             :                                         }
    1244             : 
    1245           0 :                                         vaddr += VALUE_2MB;
    1246           0 :                                         len -= VALUE_2MB;
    1247             :                                 }
    1248             :                         }
    1249           0 :                 } else {
    1250             :                         /* This is an address managed by DPDK. Just setup the translations. */
    1251           0 :                         while (len > 0) {
    1252           0 :                                 paddr = vtophys_get_paddr_memseg((uint64_t)vaddr);
    1253           0 :                                 if (paddr == SPDK_VTOPHYS_ERROR) {
    1254           0 :                                         DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
    1255           0 :                                         return -EFAULT;
    1256             :                                 }
    1257             : 
    1258           0 :                                 rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, VALUE_2MB, paddr);
    1259           0 :                                 if (rc != 0) {
    1260           0 :                                         return rc;
    1261             :                                 }
    1262             : 
    1263           0 :                                 vaddr += VALUE_2MB;
    1264           0 :                                 len -= VALUE_2MB;
    1265             :                         }
    1266             :                 }
    1267             : 
    1268           0 :                 break;
    1269             :         case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
    1270             : #if VFIO_ENABLED
    1271             :                 if (paddr == SPDK_VTOPHYS_ERROR) {
    1272             :                         /*
    1273             :                          * This is not an address that DPDK is managing.
    1274             :                          */
    1275             : 
    1276             :                         /* Check if this is a PCI BAR. They need special handling */
    1277             :                         paddr = vtophys_get_paddr_pci((uint64_t)vaddr, len);
    1278             :                         if (paddr != SPDK_VTOPHYS_ERROR) {
    1279             :                                 /* Get paddr for each 2MB chunk in this address range */
    1280             :                                 while (len > 0) {
    1281             :                                         paddr = vtophys_get_paddr_pci((uint64_t)vaddr, VALUE_2MB);
    1282             :                                         if (paddr == SPDK_VTOPHYS_ERROR) {
    1283             :                                                 DEBUG_PRINT("could not get phys addr for %p\n", vaddr);
    1284             :                                                 return -EFAULT;
    1285             :                                         }
    1286             : 
    1287             :                                         rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
    1288             :                                         if (rc != 0) {
    1289             :                                                 return rc;
    1290             :                                         }
    1291             : 
    1292             :                                         vaddr += VALUE_2MB;
    1293             :                                         len -= VALUE_2MB;
    1294             :                                 }
    1295             : 
    1296             :                                 return 0;
    1297             :                         }
    1298             : 
    1299             :                         /* If vfio is enabled,
    1300             :                          * we need to unmap the range from the IOMMU
    1301             :                          */
    1302             :                         if (spdk_iommu_is_enabled()) {
    1303             :                                 uint64_t buffer_len = len;
    1304             :                                 uint8_t *va = vaddr;
    1305             :                                 enum rte_iova_mode iova_mode;
    1306             : 
    1307             :                                 iova_mode = rte_eal_iova_mode();
    1308             :                                 /*
    1309             :                                  * In virtual address mode, the region is contiguous and can be done in
    1310             :                                  * one unmap.
    1311             :                                  */
    1312             :                                 if (iova_mode == RTE_IOVA_VA) {
    1313             :                                         paddr = spdk_mem_map_translate(map, (uint64_t)va, &buffer_len);
    1314             :                                         if (buffer_len != len || paddr != (uintptr_t)va) {
    1315             :                                                 DEBUG_PRINT("Unmapping %p with length %lu failed because "
    1316             :                                                             "translation had address 0x%" PRIx64 " and length %lu\n",
    1317             :                                                             va, len, paddr, buffer_len);
    1318             :                                                 return -EINVAL;
    1319             :                                         }
    1320             :                                         rc = vtophys_iommu_unmap_dma(paddr, len);
    1321             :                                         if (rc) {
    1322             :                                                 DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
    1323             :                                                 return -EFAULT;
    1324             :                                         }
    1325             :                                 } else if (iova_mode == RTE_IOVA_PA) {
    1326             :                                         /* Get paddr for each 2MB chunk in this address range */
    1327             :                                         while (buffer_len > 0) {
    1328             :                                                 paddr = spdk_mem_map_translate(map, (uint64_t)va, NULL);
    1329             : 
    1330             :                                                 if (paddr == SPDK_VTOPHYS_ERROR || buffer_len < VALUE_2MB) {
    1331             :                                                         DEBUG_PRINT("could not get phys addr for %p\n", va);
    1332             :                                                         return -EFAULT;
    1333             :                                                 }
    1334             : 
    1335             :                                                 rc = vtophys_iommu_unmap_dma(paddr, VALUE_2MB);
    1336             :                                                 if (rc) {
    1337             :                                                         DEBUG_PRINT("Failed to iommu unmap paddr 0x%" PRIx64 "\n", paddr);
    1338             :                                                         return -EFAULT;
    1339             :                                                 }
    1340             : 
    1341             :                                                 va += VALUE_2MB;
    1342             :                                                 buffer_len -= VALUE_2MB;
    1343             :                                         }
    1344             :                                 }
    1345             :                         }
    1346             :                 }
    1347             : #endif
    1348           0 :                 while (len > 0) {
    1349           0 :                         rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, VALUE_2MB);
    1350           0 :                         if (rc != 0) {
    1351           0 :                                 return rc;
    1352             :                         }
    1353             : 
    1354           0 :                         vaddr += VALUE_2MB;
    1355           0 :                         len -= VALUE_2MB;
    1356             :                 }
    1357             : 
    1358           0 :                 break;
    1359             :         default:
    1360           0 :                 SPDK_UNREACHABLE();
    1361             :         }
    1362             : 
    1363           0 :         return rc;
    1364           0 : }
    1365             : 
    1366             : static int
    1367           0 : numa_notify(void *cb_ctx, struct spdk_mem_map *map,
    1368             :             enum spdk_mem_map_notify_action action,
    1369             :             void *vaddr, size_t len)
    1370             : {
    1371             :         struct rte_memseg *seg;
    1372             : 
    1373             :         /* We always return 0 from here, even if we aren't able to get a
    1374             :          * memseg for the address. This can happen in non-DPDK memory
    1375             :          * registration paths, for example vhost or vfio-user. That is OK,
    1376             :          * spdk_mem_get_numa_id() just returns SPDK_ENV_NUMA_ID_ANY for
    1377             :          * that kind of memory. If we return an error here, the
    1378             :          * spdk_mem_register() from vhost or vfio-user would fail which is
    1379             :          * not what we want.
    1380             :          */
    1381           0 :         seg = rte_mem_virt2memseg(vaddr, NULL);
    1382           0 :         if (seg == NULL) {
    1383           0 :                 return 0;
    1384             :         }
    1385             : 
    1386           0 :         switch (action) {
    1387             :         case SPDK_MEM_MAP_NOTIFY_REGISTER:
    1388           0 :                 spdk_mem_map_set_translation(map, (uint64_t)vaddr, len, seg->socket_id);
    1389           0 :                 break;
    1390             :         case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
    1391           0 :                 spdk_mem_map_clear_translation(map, (uint64_t)vaddr, len);
    1392           0 :                 break;
    1393             :         default:
    1394           0 :                 break;
    1395             :         }
    1396             : 
    1397           0 :         return 0;
    1398           0 : }
    1399             : 
    1400             : static int
    1401           0 : vtophys_check_contiguous_entries(uint64_t paddr1, uint64_t paddr2)
    1402             : {
    1403             :         /* This function is always called with paddrs for two subsequent
    1404             :          * 2MB chunks in virtual address space, so those chunks will be only
    1405             :          * physically contiguous if the physical addresses are 2MB apart
    1406             :          * from each other as well.
    1407             :          */
    1408           0 :         return (paddr2 - paddr1 == VALUE_2MB);
    1409             : }
    1410             : 
    1411             : #if VFIO_ENABLED
    1412             : 
    1413             : static bool
    1414             : vfio_enabled(void)
    1415             : {
    1416             :         return rte_vfio_is_enabled("vfio_pci");
    1417             : }
    1418             : 
    1419             : /* Check if IOMMU is enabled on the system */
    1420             : static bool
    1421             : has_iommu_groups(void)
    1422             : {
    1423             :         int count = 0;
    1424             :         DIR *dir = opendir("/sys/kernel/iommu_groups");
    1425             : 
    1426             :         if (dir == NULL) {
    1427             :                 return false;
    1428             :         }
    1429             : 
    1430             :         while (count < 3 && readdir(dir) != NULL) {
    1431             :                 count++;
    1432             :         }
    1433             : 
    1434             :         closedir(dir);
    1435             :         /* there will always be ./ and ../ entries */
    1436             :         return count > 2;
    1437             : }
    1438             : 
    1439             : static bool
    1440             : vfio_noiommu_enabled(void)
    1441             : {
    1442             :         return rte_vfio_noiommu_is_enabled();
    1443             : }
    1444             : 
    1445             : static void
    1446             : vtophys_iommu_init(void)
    1447             : {
    1448             :         char proc_fd_path[PATH_MAX + 1];
    1449             :         char link_path[PATH_MAX + 1];
    1450             :         const char vfio_path[] = "/dev/vfio/vfio";
    1451             :         DIR *dir;
    1452             :         struct dirent *d;
    1453             : 
    1454             :         if (!vfio_enabled()) {
    1455             :                 return;
    1456             :         }
    1457             : 
    1458             :         if (vfio_noiommu_enabled()) {
    1459             :                 g_vfio.noiommu_enabled = true;
    1460             :         } else if (!has_iommu_groups()) {
    1461             :                 return;
    1462             :         }
    1463             : 
    1464             :         dir = opendir("/proc/self/fd");
    1465             :         if (!dir) {
    1466             :                 DEBUG_PRINT("Failed to open /proc/self/fd (%d)\n", errno);
    1467             :                 return;
    1468             :         }
    1469             : 
    1470             :         while ((d = readdir(dir)) != NULL) {
    1471             :                 if (d->d_type != DT_LNK) {
    1472             :                         continue;
    1473             :                 }
    1474             : 
    1475             :                 snprintf(proc_fd_path, sizeof(proc_fd_path), "/proc/self/fd/%s", d->d_name);
    1476             :                 if (readlink(proc_fd_path, link_path, sizeof(link_path)) != (sizeof(vfio_path) - 1)) {
    1477             :                         continue;
    1478             :                 }
    1479             : 
    1480             :                 if (memcmp(link_path, vfio_path, sizeof(vfio_path) - 1) == 0) {
    1481             :                         sscanf(d->d_name, "%d", &g_vfio.fd);
    1482             :                         break;
    1483             :                 }
    1484             :         }
    1485             : 
    1486             :         closedir(dir);
    1487             : 
    1488             :         if (g_vfio.fd < 0) {
    1489             :                 DEBUG_PRINT("Failed to discover DPDK VFIO container fd.\n");
    1490             :                 return;
    1491             :         }
    1492             : 
    1493             :         g_vfio.enabled = true;
    1494             : 
    1495             :         return;
    1496             : }
    1497             : 
    1498             : #endif
    1499             : 
    1500             : void
    1501           0 : vtophys_pci_device_added(struct rte_pci_device *pci_device)
    1502             : {
    1503             :         struct spdk_vtophys_pci_device *vtophys_dev;
    1504             : 
    1505           0 :         pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
    1506             : 
    1507           0 :         vtophys_dev = calloc(1, sizeof(*vtophys_dev));
    1508           0 :         if (vtophys_dev) {
    1509           0 :                 vtophys_dev->pci_device = pci_device;
    1510           0 :                 TAILQ_INSERT_TAIL(&g_vtophys_pci_devices, vtophys_dev, tailq);
    1511           0 :         } else {
    1512           0 :                 DEBUG_PRINT("Memory allocation error\n");
    1513             :         }
    1514           0 :         pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
    1515             : 
    1516             : #if VFIO_ENABLED
    1517             :         struct spdk_vfio_dma_map *dma_map;
    1518             :         int ret;
    1519             : 
    1520             :         if (!g_vfio.enabled) {
    1521             :                 return;
    1522             :         }
    1523             : 
    1524             :         pthread_mutex_lock(&g_vfio.mutex);
    1525             :         g_vfio.device_ref++;
    1526             :         if (g_vfio.device_ref > 1) {
    1527             :                 pthread_mutex_unlock(&g_vfio.mutex);
    1528             :                 return;
    1529             :         }
    1530             : 
    1531             :         /* This is the first SPDK device using DPDK vfio. This means that the first
    1532             :          * IOMMU group might have been just been added to the DPDK vfio container.
    1533             :          * From this point it is certain that the memory can be mapped now.
    1534             :          */
    1535             :         TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
    1536             :                 ret = ioctl(g_vfio.fd, VFIO_IOMMU_MAP_DMA, &dma_map->map);
    1537             :                 if (ret) {
    1538             :                         DEBUG_PRINT("Cannot update DMA mapping, error %d\n", errno);
    1539             :                         break;
    1540             :                 }
    1541             :         }
    1542             :         pthread_mutex_unlock(&g_vfio.mutex);
    1543             : #endif
    1544           0 : }
    1545             : 
    1546             : void
    1547           0 : vtophys_pci_device_removed(struct rte_pci_device *pci_device)
    1548             : {
    1549             :         struct spdk_vtophys_pci_device *vtophys_dev;
    1550             : 
    1551           0 :         pthread_mutex_lock(&g_vtophys_pci_devices_mutex);
    1552           0 :         TAILQ_FOREACH(vtophys_dev, &g_vtophys_pci_devices, tailq) {
    1553           0 :                 if (vtophys_dev->pci_device == pci_device) {
    1554           0 :                         TAILQ_REMOVE(&g_vtophys_pci_devices, vtophys_dev, tailq);
    1555           0 :                         free(vtophys_dev);
    1556           0 :                         break;
    1557             :                 }
    1558           0 :         }
    1559           0 :         pthread_mutex_unlock(&g_vtophys_pci_devices_mutex);
    1560             : 
    1561             : #if VFIO_ENABLED
    1562             :         struct spdk_vfio_dma_map *dma_map;
    1563             :         int ret;
    1564             : 
    1565             :         if (!g_vfio.enabled) {
    1566             :                 return;
    1567             :         }
    1568             : 
    1569             :         pthread_mutex_lock(&g_vfio.mutex);
    1570             :         assert(g_vfio.device_ref > 0);
    1571             :         g_vfio.device_ref--;
    1572             :         if (g_vfio.device_ref > 0) {
    1573             :                 pthread_mutex_unlock(&g_vfio.mutex);
    1574             :                 return;
    1575             :         }
    1576             : 
    1577             :         /* This is the last SPDK device using DPDK vfio. If DPDK doesn't have
    1578             :          * any additional devices using it's vfio container, all the mappings
    1579             :          * will be automatically removed by the Linux vfio driver. We unmap
    1580             :          * the memory manually to be able to easily re-map it later regardless
    1581             :          * of other, external factors.
    1582             :          */
    1583             :         TAILQ_FOREACH(dma_map, &g_vfio.maps, tailq) {
    1584             :                 struct vfio_iommu_type1_dma_unmap unmap = {};
    1585             :                 unmap.argsz = sizeof(unmap);
    1586             :                 unmap.flags = 0;
    1587             :                 unmap.iova = dma_map->map.iova;
    1588             :                 unmap.size = dma_map->map.size;
    1589             :                 ret = ioctl(g_vfio.fd, VFIO_IOMMU_UNMAP_DMA, &unmap);
    1590             :                 if (ret) {
    1591             :                         DEBUG_PRINT("Cannot unmap DMA memory, error %d\n", errno);
    1592             :                         break;
    1593             :                 }
    1594             :         }
    1595             :         pthread_mutex_unlock(&g_vfio.mutex);
    1596             : #endif
    1597           0 : }
    1598             : 
    1599             : int
    1600           0 : vtophys_init(void)
    1601             : {
    1602           0 :         const struct spdk_mem_map_ops vtophys_map_ops = {
    1603             :                 .notify_cb = vtophys_notify,
    1604             :                 .are_contiguous = vtophys_check_contiguous_entries,
    1605             :         };
    1606             : 
    1607           0 :         const struct spdk_mem_map_ops phys_ref_map_ops = {
    1608             :                 .notify_cb = NULL,
    1609             :                 .are_contiguous = NULL,
    1610             :         };
    1611             : 
    1612           0 :         const struct spdk_mem_map_ops numa_map_ops = {
    1613             :                 .notify_cb = numa_notify,
    1614             :                 .are_contiguous = NULL,
    1615             :         };
    1616             : 
    1617             : #if VFIO_ENABLED
    1618             :         vtophys_iommu_init();
    1619             : #endif
    1620             : 
    1621           0 :         g_phys_ref_map = spdk_mem_map_alloc(0, &phys_ref_map_ops, NULL);
    1622           0 :         if (g_phys_ref_map == NULL) {
    1623           0 :                 DEBUG_PRINT("phys_ref map allocation failed.\n");
    1624           0 :                 return -ENOMEM;
    1625             :         }
    1626             : 
    1627           0 :         g_numa_map = spdk_mem_map_alloc(SPDK_ENV_NUMA_ID_ANY, &numa_map_ops, NULL);
    1628           0 :         if (g_numa_map == NULL) {
    1629           0 :                 DEBUG_PRINT("numa map allocation failed.\n");
    1630           0 :                 spdk_mem_map_free(&g_phys_ref_map);
    1631           0 :                 return -ENOMEM;
    1632             :         }
    1633             : 
    1634           0 :         if (g_huge_pages) {
    1635           0 :                 g_vtophys_map = spdk_mem_map_alloc(SPDK_VTOPHYS_ERROR, &vtophys_map_ops, NULL);
    1636           0 :                 if (g_vtophys_map == NULL) {
    1637           0 :                         DEBUG_PRINT("vtophys map allocation failed\n");
    1638           0 :                         spdk_mem_map_free(&g_numa_map);
    1639           0 :                         spdk_mem_map_free(&g_phys_ref_map);
    1640           0 :                         return -ENOMEM;
    1641             :                 }
    1642           0 :         }
    1643           0 :         return 0;
    1644           0 : }
    1645             : 
    1646             : void
    1647           0 : vtophys_fini(void)
    1648             : {
    1649           0 :         spdk_mem_map_free(&g_vtophys_map);
    1650           0 :         spdk_mem_map_free(&g_numa_map);
    1651           0 :         spdk_mem_map_free(&g_phys_ref_map);
    1652           0 : }
    1653             : 
    1654             : uint64_t
    1655           0 : spdk_vtophys(const void *buf, uint64_t *size)
    1656             : {
    1657             :         uint64_t vaddr, paddr_2mb;
    1658             : 
    1659           0 :         if (!g_huge_pages) {
    1660           0 :                 return SPDK_VTOPHYS_ERROR;
    1661             :         }
    1662             : 
    1663           0 :         vaddr = (uint64_t)buf;
    1664           0 :         paddr_2mb = spdk_mem_map_translate(g_vtophys_map, vaddr, size);
    1665             : 
    1666             :         /*
    1667             :          * SPDK_VTOPHYS_ERROR has all bits set, so if the lookup returned SPDK_VTOPHYS_ERROR,
    1668             :          * we will still bitwise-or it with the buf offset below, but the result will still be
    1669             :          * SPDK_VTOPHYS_ERROR. However now that we do + rather than | (due to PCI vtophys being
    1670             :          * unaligned) we must now check the return value before addition.
    1671             :          */
    1672             :         SPDK_STATIC_ASSERT(SPDK_VTOPHYS_ERROR == UINT64_C(-1), "SPDK_VTOPHYS_ERROR should be all 1s");
    1673           0 :         if (paddr_2mb == SPDK_VTOPHYS_ERROR) {
    1674           0 :                 return SPDK_VTOPHYS_ERROR;
    1675             :         } else {
    1676           0 :                 return paddr_2mb + (vaddr & MASK_2MB);
    1677             :         }
    1678           0 : }
    1679             : 
    1680             : int32_t
    1681           0 : spdk_mem_get_numa_id(const void *buf, uint64_t *size)
    1682             : {
    1683           0 :         return spdk_mem_map_translate(g_numa_map, (uint64_t)buf, size);
    1684             : }
    1685             : 
    1686             : int
    1687           0 : spdk_mem_get_fd_and_offset(void *vaddr, uint64_t *offset)
    1688             : {
    1689             :         struct rte_memseg *seg;
    1690             :         int ret, fd;
    1691             : 
    1692           0 :         seg = rte_mem_virt2memseg(vaddr, NULL);
    1693           0 :         if (!seg) {
    1694           0 :                 SPDK_ERRLOG("memory %p doesn't exist\n", vaddr);
    1695           0 :                 return -ENOENT;
    1696             :         }
    1697             : 
    1698           0 :         fd = rte_memseg_get_fd_thread_unsafe(seg);
    1699           0 :         if (fd < 0) {
    1700           0 :                 return fd;
    1701             :         }
    1702             : 
    1703           0 :         ret = rte_memseg_get_fd_offset_thread_unsafe(seg, offset);
    1704           0 :         if (ret < 0) {
    1705           0 :                 return ret;
    1706             :         }
    1707             : 
    1708           0 :         return fd;
    1709           0 : }
    1710             : 
    1711             : void
    1712           0 : mem_disable_huge_pages(void)
    1713             : {
    1714           0 :         g_huge_pages = false;
    1715           0 :         mem_map_use_page_shift(SHIFT_4KB);
    1716           0 : }
    1717             : 
    1718             : void
    1719           0 : mem_map_use_page_shift(uint32_t page_shift)
    1720             : {
    1721           0 :         g_map_page_cfg.shift = page_shift;
    1722           0 :         g_map_page_cfg.size = 1UL << page_shift;
    1723           0 :         g_map_page_cfg.mask = g_map_page_cfg.size - 1;
    1724           0 :         g_map_page_cfg.num_pages_per_gb = 1UL << (SHIFT_1GB - page_shift);
    1725           0 : }

Generated by: LCOV version 1.15