* Copyright 2019 Mellanox Technologies, Ltd
*/
-#include <dlfcn.h>
#include <unistd.h>
#include <string.h>
+#include <stdio.h>
#include <rte_errno.h>
+#include <rte_mempool.h>
+#include <rte_malloc.h>
#include "mlx5_common.h"
+#include "mlx5_common_os.h"
#include "mlx5_common_utils.h"
-#include "mlx5_glue.h"
-
+#include "mlx5_malloc.h"
+#include "mlx5_common_pci.h"
int mlx5_common_logtype;
+uint8_t haswell_broadwell_cpu;
+
+/* In case this is an x86_64 intel processor to check if
+ * we should use relaxed ordering.
+ */
+#ifdef RTE_ARCH_X86_64
+/**
+ * This function returns processor identification and feature information
+ * into the registers.
+ *
+ * @param eax, ebx, ecx, edx
+ * Pointers to the registers that will hold cpu information.
+ * @param level
+ * The main category of information returned.
+ */
+static inline void mlx5_cpu_id(unsigned int level,
+ unsigned int *eax, unsigned int *ebx,
+ unsigned int *ecx, unsigned int *edx)
+{
+ __asm__("cpuid\n\t"
+ : "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx)
+ : "0" (level));
+}
+#endif
+
+RTE_INIT_PRIO(mlx5_log_init, LOG)
+{
+ mlx5_common_logtype = rte_log_register("pmd.common.mlx5");
+ if (mlx5_common_logtype >= 0)
+ rte_log_set_level(mlx5_common_logtype, RTE_LOG_NOTICE);
+}
+
+static bool mlx5_common_initialized;
+
+/**
+ * One time innitialization routine for run-time dependency on glue library
+ * for multiple PMDs. Each mlx5 PMD that depends on mlx5_common module,
+ * must invoke in its constructor.
+ */
+void
+mlx5_common_init(void)
+{
+ if (mlx5_common_initialized)
+ return;
+
+ mlx5_glue_constructor();
+ mlx5_common_pci_init();
+ mlx5_common_initialized = true;
+}
+
+/**
+ * This function is responsible of initializing the variable
+ * haswell_broadwell_cpu by checking if the cpu is intel
+ * and reading the data returned from mlx5_cpu_id().
+ * since haswell and broadwell cpus don't have improved performance
+ * when using relaxed ordering we want to check the cpu type before
+ * before deciding whether to enable RO or not.
+ * if the cpu is haswell or broadwell the variable will be set to 1
+ * otherwise it will be 0.
+ */
+RTE_INIT_PRIO(mlx5_is_haswell_broadwell_cpu, LOG)
+{
+#ifdef RTE_ARCH_X86_64
+ unsigned int broadwell_models[4] = {0x3d, 0x47, 0x4F, 0x56};
+ unsigned int haswell_models[4] = {0x3c, 0x3f, 0x45, 0x46};
+ unsigned int i, model, family, brand_id, vendor;
+ unsigned int signature_intel_ebx = 0x756e6547;
+ unsigned int extended_model;
+ unsigned int eax = 0;
+ unsigned int ebx = 0;
+ unsigned int ecx = 0;
+ unsigned int edx = 0;
+ int max_level;
+
+ mlx5_cpu_id(0, &eax, &ebx, &ecx, &edx);
+ vendor = ebx;
+ max_level = eax;
+ if (max_level < 1) {
+ haswell_broadwell_cpu = 0;
+ return;
+ }
+ mlx5_cpu_id(1, &eax, &ebx, &ecx, &edx);
+ model = (eax >> 4) & 0x0f;
+ family = (eax >> 8) & 0x0f;
+ brand_id = ebx & 0xff;
+ extended_model = (eax >> 12) & 0xf0;
+ /* Check if the processor is Haswell or Broadwell */
+ if (vendor == signature_intel_ebx) {
+ if (family == 0x06)
+ model += extended_model;
+ if (brand_id == 0 && family == 0x6) {
+ for (i = 0; i < RTE_DIM(broadwell_models); i++)
+ if (model == broadwell_models[i]) {
+ haswell_broadwell_cpu = 1;
+ return;
+ }
+ for (i = 0; i < RTE_DIM(haswell_models); i++)
+ if (model == haswell_models[i]) {
+ haswell_broadwell_cpu = 1;
+ return;
+ }
+ }
+ }
+#endif
+ haswell_broadwell_cpu = 0;
+}
+
+/**
+ * Allocate page of door-bells and register it using DevX API.
+ *
+ * @param [in] ctx
+ * Pointer to the device context.
+ *
+ * @return
+ * Pointer to new page on success, NULL otherwise.
+ */
+static struct mlx5_devx_dbr_page *
+mlx5_alloc_dbr_page(void *ctx)
+{
+ struct mlx5_devx_dbr_page *page;
-#ifdef RTE_IBVERBS_LINK_DLOPEN
+ /* Allocate space for door-bell page and management data. */
+ page = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
+ sizeof(struct mlx5_devx_dbr_page),
+ RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
+ if (!page) {
+ DRV_LOG(ERR, "cannot allocate dbr page");
+ return NULL;
+ }
+ /* Register allocated memory. */
+ page->umem = mlx5_os_umem_reg(ctx, page->dbrs,
+ MLX5_DBR_PAGE_SIZE, 0);
+ if (!page->umem) {
+ DRV_LOG(ERR, "cannot umem reg dbr page");
+ mlx5_free(page);
+ return NULL;
+ }
+ return page;
+}
/**
- * Suffix RTE_EAL_PMD_PATH with "-glue".
+ * Find the next available door-bell, allocate new page if needed.
+ *
+ * @param [in] ctx
+ * Pointer to device context.
+ * @param [in] head
+ * Pointer to the head of dbr pages list.
+ * @param [out] dbr_page
+ * Door-bell page containing the page data.
*
- * This function performs a sanity check on RTE_EAL_PMD_PATH before
- * suffixing its last component.
+ * @return
+ * Door-bell address offset on success, a negative error value otherwise.
+ */
+int64_t
+mlx5_get_dbr(void *ctx, struct mlx5_dbr_page_list *head,
+ struct mlx5_devx_dbr_page **dbr_page)
+{
+ struct mlx5_devx_dbr_page *page = NULL;
+ uint32_t i, j;
+
+ LIST_FOREACH(page, head, next)
+ if (page->dbr_count < MLX5_DBR_PER_PAGE)
+ break;
+ if (!page) { /* No page with free door-bell exists. */
+ page = mlx5_alloc_dbr_page(ctx);
+ if (!page) /* Failed to allocate new page. */
+ return (-1);
+ LIST_INSERT_HEAD(head, page, next);
+ }
+ /* Loop to find bitmap part with clear bit. */
+ for (i = 0;
+ i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
+ i++)
+ ; /* Empty. */
+ /* Find the first clear bit. */
+ MLX5_ASSERT(i < MLX5_DBR_BITMAP_SIZE);
+ j = rte_bsf64(~page->dbr_bitmap[i]);
+ page->dbr_bitmap[i] |= (UINT64_C(1) << j);
+ page->dbr_count++;
+ *dbr_page = page;
+ return (i * CHAR_BIT * sizeof(uint64_t) + j) * MLX5_DBR_SIZE;
+}
+
+/**
+ * Release a door-bell record.
*
- * @param buf[out]
- * Output buffer, should be large enough otherwise NULL is returned.
- * @param size
- * Size of @p out.
+ * @param [in] head
+ * Pointer to the head of dbr pages list.
+ * @param [in] umem_id
+ * UMEM ID of page containing the door-bell record to release.
+ * @param [in] offset
+ * Offset of door-bell record in page.
*
* @return
- * Pointer to @p buf or @p NULL in case suffix cannot be appended.
+ * 0 on success, a negative error value otherwise.
*/
-static char *
-mlx5_glue_path(char *buf, size_t size)
+int32_t
+mlx5_release_dbr(struct mlx5_dbr_page_list *head, uint32_t umem_id,
+ uint64_t offset)
{
- static const char *const bad[] = { "/", ".", "..", NULL };
- const char *path = RTE_EAL_PMD_PATH;
- size_t len = strlen(path);
- size_t off;
- int i;
-
- while (len && path[len - 1] == '/')
- --len;
- for (off = len; off && path[off - 1] != '/'; --off)
- ;
- for (i = 0; bad[i]; ++i)
- if (!strncmp(path + off, bad[i], (int)(len - off)))
- goto error;
- i = snprintf(buf, size, "%.*s-glue", (int)len, path);
- if (i == -1 || (size_t)i >= size)
- goto error;
- return buf;
-error:
- RTE_LOG(ERR, PMD, "unable to append \"-glue\" to last component of"
- " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"), please"
- " re-configure DPDK");
- return NULL;
+ struct mlx5_devx_dbr_page *page = NULL;
+ int ret = 0;
+
+ LIST_FOREACH(page, head, next)
+ /* Find the page this address belongs to. */
+ if (mlx5_os_get_umem_id(page->umem) == umem_id)
+ break;
+ if (!page)
+ return -EINVAL;
+ page->dbr_count--;
+ if (!page->dbr_count) {
+ /* Page not used, free it and remove from list. */
+ LIST_REMOVE(page, next);
+ if (page->umem)
+ ret = -mlx5_os_umem_dereg(page->umem);
+ mlx5_free(page);
+ } else {
+ /* Mark in bitmap that this door-bell is not in use. */
+ offset /= MLX5_DBR_SIZE;
+ int i = offset / 64;
+ int j = offset % 64;
+
+ page->dbr_bitmap[i] &= ~(UINT64_C(1) << j);
+ }
+ return ret;
}
-#endif
/**
- * Initialization routine for run-time dependency on rdma-core.
+ * Allocate the User Access Region with DevX on specified device.
+ *
+ * @param [in] ctx
+ * Infiniband device context to perform allocation on.
+ * @param [in] mapping
+ * MLX5DV_UAR_ALLOC_TYPE_BF - allocate as cached memory with write-combining
+ * attributes (if supported by the host), the
+ * writes to the UAR registers must be followed
+ * by write memory barrier.
+ * MLX5DV_UAR_ALLOC_TYPE_NC - allocate as non-cached nenory, all writes are
+ * promoted to the registers immediately, no
+ * memory barriers needed.
+ * mapping < 0 - the first attempt is performed with MLX5DV_UAR_ALLOC_TYPE_BF,
+ * if this fails the next attempt with MLX5DV_UAR_ALLOC_TYPE_NC
+ * is performed. The drivers specifying negative values should
+ * always provide the write memory barrier operation after UAR
+ * register writings.
+ * If there is no definitions for the MLX5DV_UAR_ALLOC_TYPE_xx (older rdma
+ * library headers), the caller can specify 0.
+ *
+ * @return
+ * UAR object pointer on success, NULL otherwise and rte_errno is set.
*/
-RTE_INIT_PRIO(mlx5_glue_init, CLASS)
+void *
+mlx5_devx_alloc_uar(void *ctx, int mapping)
{
- void *handle = NULL;
+ void *uar;
+ uint32_t retry, uar_mapping;
+ void *base_addr;
- /* Initialize common log type. */
- mlx5_common_logtype = rte_log_register("pmd.common.mlx5");
- if (mlx5_common_logtype >= 0)
- rte_log_set_level(mlx5_common_logtype, RTE_LOG_NOTICE);
- /*
- * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
- * huge pages. Calling ibv_fork_init() during init allows
- * applications to use fork() safely for purposes other than
- * using this PMD, which is not supported in forked processes.
- */
- setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
- /* Match the size of Rx completion entry to the size of a cacheline. */
- if (RTE_CACHE_LINE_SIZE == 128)
- setenv("MLX5_CQE_SIZE", "128", 0);
- /*
- * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
- * cleanup all the Verbs resources even when the device was removed.
- */
- setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
- /* The glue initialization was done earlier by mlx5 common library. */
-#ifdef RTE_IBVERBS_LINK_DLOPEN
- char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
- const char *path[] = {
+ for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
+#ifdef MLX5DV_UAR_ALLOC_TYPE_NC
+ /* Control the mapping type according to the settings. */
+ uar_mapping = (mapping < 0) ?
+ MLX5DV_UAR_ALLOC_TYPE_NC : mapping;
+#else
/*
- * A basic security check is necessary before trusting
- * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
+ * It seems we have no way to control the memory mapping type
+ * for the UAR, the default "Write-Combining" type is supposed.
*/
- (geteuid() == getuid() && getegid() == getgid() ?
- getenv("MLX5_GLUE_PATH") : NULL),
- /*
- * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
- * variant, otherwise let dlopen() look up libraries on its
- * own.
- */
- (*RTE_EAL_PMD_PATH ?
- mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
- };
- unsigned int i = 0;
- void **sym;
- const char *dlmsg;
-
- while (!handle && i != RTE_DIM(path)) {
- const char *end;
- size_t len;
- int ret;
-
- if (!path[i]) {
- ++i;
- continue;
+ uar_mapping = 0;
+ RTE_SET_USED(mapping);
+#endif
+ uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
+#ifdef MLX5DV_UAR_ALLOC_TYPE_NC
+ if (!uar &&
+ mapping < 0 &&
+ uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
+ /*
+ * In some environments like virtual machine the
+ * Write Combining mapped might be not supported and
+ * UAR allocation fails. We tried "Non-Cached" mapping
+ * for the case.
+ */
+ DRV_LOG(WARNING, "Failed to allocate DevX UAR (BF)");
+ uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
+ uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
+ } else if (!uar &&
+ mapping < 0 &&
+ uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
+ /*
+ * If Verbs/kernel does not support "Non-Cached"
+ * try the "Write-Combining".
+ */
+ DRV_LOG(WARNING, "Failed to allocate DevX UAR (NC)");
+ uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
+ uar = mlx5_glue->devx_alloc_uar(ctx, uar_mapping);
}
- end = strpbrk(path[i], ":;");
- if (!end)
- end = path[i] + strlen(path[i]);
- len = end - path[i];
- ret = 0;
- do {
- char name[ret + 1];
-
- ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
- (int)len, path[i],
- (!len || *(end - 1) == '/') ? "" : "/");
- if (ret == -1)
- break;
- if (sizeof(name) != (size_t)ret + 1)
- continue;
- DRV_LOG(DEBUG, "Looking for rdma-core glue as "
- "\"%s\"", name);
- handle = dlopen(name, RTLD_LAZY);
+#endif
+ if (!uar) {
+ DRV_LOG(ERR, "Failed to allocate DevX UAR (BF/NC)");
+ rte_errno = ENOMEM;
+ goto exit;
+ }
+ base_addr = mlx5_os_get_devx_uar_base_addr(uar);
+ if (base_addr)
break;
- } while (1);
- path[i] = end + 1;
- if (!*end)
- ++i;
- }
- if (!handle) {
- rte_errno = EINVAL;
- dlmsg = dlerror();
- if (dlmsg)
- DRV_LOG(WARNING, "Cannot load glue library: %s", dlmsg);
- goto glue_error;
- }
- sym = dlsym(handle, "mlx5_glue");
- if (!sym || !*sym) {
- rte_errno = EINVAL;
- dlmsg = dlerror();
- if (dlmsg)
- DRV_LOG(ERR, "Cannot resolve glue symbol: %s", dlmsg);
- goto glue_error;
- }
- mlx5_glue = *sym;
-#endif /* RTE_IBVERBS_LINK_DLOPEN */
-#ifndef NDEBUG
- /* Glue structure must not contain any NULL pointers. */
- {
- unsigned int i;
-
- for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
- assert(((const void *const *)mlx5_glue)[i]);
+ /*
+ * The UARs are allocated by rdma_core within the
+ * IB device context, on context closure all UARs
+ * will be freed, should be no memory/object leakage.
+ */
+ DRV_LOG(WARNING, "Retrying to allocate DevX UAR");
+ uar = NULL;
}
-#endif
- if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
- rte_errno = EINVAL;
- DRV_LOG(ERR, "rdma-core glue \"%s\" mismatch: \"%s\" is "
- "required", mlx5_glue->version, MLX5_GLUE_VERSION);
- goto glue_error;
+ /* Check whether we finally succeeded with valid UAR allocation. */
+ if (!uar) {
+ DRV_LOG(ERR, "Failed to allocate DevX UAR (NULL base)");
+ rte_errno = ENOMEM;
}
- mlx5_glue->fork_init();
- return;
-glue_error:
- if (handle)
- dlclose(handle);
- DRV_LOG(WARNING, "Cannot initialize MLX5 common due to missing"
- " run-time dependency on rdma-core libraries (libibverbs,"
- " libmlx5)");
- mlx5_glue = NULL;
- return;
+ /*
+ * Return void * instead of struct mlx5dv_devx_uar *
+ * is for compatibility with older rdma-core library headers.
+ */
+exit:
+ return uar;
}