RDMA/core: Remove FMR pool API
authorMax Gurtovoy <maxg@mellanox.com>
Thu, 28 May 2020 19:45:46 +0000 (16:45 -0300)
committerJason Gunthorpe <jgg@mellanox.com>
Tue, 2 Jun 2020 23:32:53 +0000 (20:32 -0300)
This ancient and unsafe method for memory registration is no longer used
by any RDMA based ULP. Remove the FMR pool API from the core driver.

Link: https://lore.kernel.org/r/4-v3-f58e6669d5d3+2cf-fmr_removal_jgg@mellanox.com
Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Documentation/driver-api/infiniband.rst
drivers/infiniband/core/Makefile
drivers/infiniband/core/fmr_pool.c [deleted file]
include/rdma/ib_fmr_pool.h [deleted file]

index 1a3116f..30e142c 100644 (file)
@@ -37,9 +37,6 @@ InfiniBand core interfaces
 .. kernel-doc:: drivers/infiniband/core/ud_header.c
     :export:
 
-.. kernel-doc:: drivers/infiniband/core/fmr_pool.c
-    :export:
-
 .. kernel-doc:: drivers/infiniband/core/umem.c
     :export:
 
index 63c1591..24cb71a 100644 (file)
@@ -8,7 +8,7 @@ obj-$(CONFIG_INFINIBAND_USER_MAD) +=    ib_umad.o
 obj-$(CONFIG_INFINIBAND_USER_ACCESS) += ib_uverbs.o $(user_access-y)
 
 ib_core-y :=                   packer.o ud_header.o verbs.o cq.o rw.o sysfs.o \
-                               device.o fmr_pool.o cache.o netlink.o \
+                               device.o cache.o netlink.o \
                                roce_gid_mgmt.o mr_pool.o addr.o sa_query.o \
                                multicast.o mad.o smi.o agent.o mad_rmpp.o \
                                nldev.o restrack.o counters.o ib_core_uverbs.o \
diff --git a/drivers/infiniband/core/fmr_pool.c b/drivers/infiniband/core/fmr_pool.c
deleted file mode 100644 (file)
index e08aec4..0000000
+++ /dev/null
@@ -1,494 +0,0 @@
-/*
- * Copyright (c) 2004 Topspin Communications.  All rights reserved.
- * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- *
- * This software is available to you under a choice of one of two
- * licenses.  You may choose to be licensed under the terms of the GNU
- * General Public License (GPL) Version 2, available from the file
- * COPYING in the main directory of this source tree, or the
- * OpenIB.org BSD license below:
- *
- *     Redistribution and use in source and binary forms, with or
- *     without modification, are permitted provided that the following
- *     conditions are met:
- *
- *      - Redistributions of source code must retain the above
- *        copyright notice, this list of conditions and the following
- *        disclaimer.
- *
- *      - Redistributions in binary form must reproduce the above
- *        copyright notice, this list of conditions and the following
- *        disclaimer in the documentation and/or other materials
- *        provided with the distribution.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
- * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
- * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
- * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
- * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
- * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#include <linux/errno.h>
-#include <linux/spinlock.h>
-#include <linux/export.h>
-#include <linux/slab.h>
-#include <linux/jhash.h>
-#include <linux/kthread.h>
-
-#include <rdma/ib_fmr_pool.h>
-
-#include "core_priv.h"
-
-#define PFX "fmr_pool: "
-
-enum {
-       IB_FMR_MAX_REMAPS = 32,
-
-       IB_FMR_HASH_BITS  = 8,
-       IB_FMR_HASH_SIZE  = 1 << IB_FMR_HASH_BITS,
-       IB_FMR_HASH_MASK  = IB_FMR_HASH_SIZE - 1
-};
-
-/*
- * If an FMR is not in use, then the list member will point to either
- * its pool's free_list (if the FMR can be mapped again; that is,
- * remap_count < pool->max_remaps) or its pool's dirty_list (if the
- * FMR needs to be unmapped before being remapped).  In either of
- * these cases it is a bug if the ref_count is not 0.  In other words,
- * if ref_count is > 0, then the list member must not be linked into
- * either free_list or dirty_list.
- *
- * The cache_node member is used to link the FMR into a cache bucket
- * (if caching is enabled).  This is independent of the reference
- * count of the FMR.  When a valid FMR is released, its ref_count is
- * decremented, and if ref_count reaches 0, the FMR is placed in
- * either free_list or dirty_list as appropriate.  However, it is not
- * removed from the cache and may be "revived" if a call to
- * ib_fmr_register_physical() occurs before the FMR is remapped.  In
- * this case we just increment the ref_count and remove the FMR from
- * free_list/dirty_list.
- *
- * Before we remap an FMR from free_list, we remove it from the cache
- * (to prevent another user from obtaining a stale FMR).  When an FMR
- * is released, we add it to the tail of the free list, so that our
- * cache eviction policy is "least recently used."
- *
- * All manipulation of ref_count, list and cache_node is protected by
- * pool_lock to maintain consistency.
- */
-
-struct ib_fmr_pool {
-       spinlock_t                pool_lock;
-
-       int                       pool_size;
-       int                       max_pages;
-       int                       max_remaps;
-       int                       dirty_watermark;
-       int                       dirty_len;
-       struct list_head          free_list;
-       struct list_head          dirty_list;
-       struct hlist_head        *cache_bucket;
-
-       void                     (*flush_function)(struct ib_fmr_pool *pool,
-                                                  void *              arg);
-       void                     *flush_arg;
-
-       struct kthread_worker     *worker;
-       struct kthread_work       work;
-
-       atomic_t                  req_ser;
-       atomic_t                  flush_ser;
-
-       wait_queue_head_t         force_wait;
-};
-
-static inline u32 ib_fmr_hash(u64 first_page)
-{
-       return jhash_2words((u32) first_page, (u32) (first_page >> 32), 0) &
-               (IB_FMR_HASH_SIZE - 1);
-}
-
-/* Caller must hold pool_lock */
-static inline struct ib_pool_fmr *ib_fmr_cache_lookup(struct ib_fmr_pool *pool,
-                                                     u64 *page_list,
-                                                     int  page_list_len,
-                                                     u64  io_virtual_address)
-{
-       struct hlist_head *bucket;
-       struct ib_pool_fmr *fmr;
-
-       if (!pool->cache_bucket)
-               return NULL;
-
-       bucket = pool->cache_bucket + ib_fmr_hash(*page_list);
-
-       hlist_for_each_entry(fmr, bucket, cache_node)
-               if (io_virtual_address == fmr->io_virtual_address &&
-                   page_list_len      == fmr->page_list_len      &&
-                   !memcmp(page_list, fmr->page_list,
-                           page_list_len * sizeof *page_list))
-                       return fmr;
-
-       return NULL;
-}
-
-static void ib_fmr_batch_release(struct ib_fmr_pool *pool)
-{
-       int                 ret;
-       struct ib_pool_fmr *fmr;
-       LIST_HEAD(unmap_list);
-       LIST_HEAD(fmr_list);
-
-       spin_lock_irq(&pool->pool_lock);
-
-       list_for_each_entry(fmr, &pool->dirty_list, list) {
-               hlist_del_init(&fmr->cache_node);
-               fmr->remap_count = 0;
-               list_add_tail(&fmr->fmr->list, &fmr_list);
-       }
-
-       list_splice_init(&pool->dirty_list, &unmap_list);
-       pool->dirty_len = 0;
-
-       spin_unlock_irq(&pool->pool_lock);
-
-       if (list_empty(&unmap_list)) {
-               return;
-       }
-
-       ret = ib_unmap_fmr(&fmr_list);
-       if (ret)
-               pr_warn(PFX "ib_unmap_fmr returned %d\n", ret);
-
-       spin_lock_irq(&pool->pool_lock);
-       list_splice(&unmap_list, &pool->free_list);
-       spin_unlock_irq(&pool->pool_lock);
-}
-
-static void ib_fmr_cleanup_func(struct kthread_work *work)
-{
-       struct ib_fmr_pool *pool = container_of(work, struct ib_fmr_pool, work);
-
-       ib_fmr_batch_release(pool);
-       atomic_inc(&pool->flush_ser);
-       wake_up_interruptible(&pool->force_wait);
-
-       if (pool->flush_function)
-               pool->flush_function(pool, pool->flush_arg);
-
-       if (atomic_read(&pool->flush_ser) - atomic_read(&pool->req_ser) < 0)
-               kthread_queue_work(pool->worker, &pool->work);
-}
-
-/**
- * ib_create_fmr_pool - Create an FMR pool
- * @pd:Protection domain for FMRs
- * @params:FMR pool parameters
- *
- * Create a pool of FMRs.  Return value is pointer to new pool or
- * error code if creation failed.
- */
-struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
-                                      struct ib_fmr_pool_param *params)
-{
-       struct ib_device   *device;
-       struct ib_fmr_pool *pool;
-       int i;
-       int ret;
-       int max_remaps;
-
-       if (!params)
-               return ERR_PTR(-EINVAL);
-
-       device = pd->device;
-       if (!device->ops.alloc_fmr    || !device->ops.dealloc_fmr  ||
-           !device->ops.map_phys_fmr || !device->ops.unmap_fmr) {
-               dev_info(&device->dev, "Device does not support FMRs\n");
-               return ERR_PTR(-ENOSYS);
-       }
-
-       if (!device->attrs.max_map_per_fmr)
-               max_remaps = IB_FMR_MAX_REMAPS;
-       else
-               max_remaps = device->attrs.max_map_per_fmr;
-
-       pool = kmalloc(sizeof *pool, GFP_KERNEL);
-       if (!pool)
-               return ERR_PTR(-ENOMEM);
-
-       pool->cache_bucket   = NULL;
-       pool->flush_function = params->flush_function;
-       pool->flush_arg      = params->flush_arg;
-
-       INIT_LIST_HEAD(&pool->free_list);
-       INIT_LIST_HEAD(&pool->dirty_list);
-
-       if (params->cache) {
-               pool->cache_bucket =
-                       kmalloc_array(IB_FMR_HASH_SIZE,
-                                     sizeof(*pool->cache_bucket),
-                                     GFP_KERNEL);
-               if (!pool->cache_bucket) {
-                       ret = -ENOMEM;
-                       goto out_free_pool;
-               }
-
-               for (i = 0; i < IB_FMR_HASH_SIZE; ++i)
-                       INIT_HLIST_HEAD(pool->cache_bucket + i);
-       }
-
-       pool->pool_size       = 0;
-       pool->max_pages       = params->max_pages_per_fmr;
-       pool->max_remaps      = max_remaps;
-       pool->dirty_watermark = params->dirty_watermark;
-       pool->dirty_len       = 0;
-       spin_lock_init(&pool->pool_lock);
-       atomic_set(&pool->req_ser,   0);
-       atomic_set(&pool->flush_ser, 0);
-       init_waitqueue_head(&pool->force_wait);
-
-       pool->worker =
-               kthread_create_worker(0, "ib_fmr(%s)", dev_name(&device->dev));
-       if (IS_ERR(pool->worker)) {
-               pr_warn(PFX "couldn't start cleanup kthread worker\n");
-               ret = PTR_ERR(pool->worker);
-               goto out_free_pool;
-       }
-       kthread_init_work(&pool->work, ib_fmr_cleanup_func);
-
-       {
-               struct ib_pool_fmr *fmr;
-               struct ib_fmr_attr fmr_attr = {
-                       .max_pages  = params->max_pages_per_fmr,
-                       .max_maps   = pool->max_remaps,
-                       .page_shift = params->page_shift
-               };
-               int bytes_per_fmr = sizeof *fmr;
-
-               if (pool->cache_bucket)
-                       bytes_per_fmr += params->max_pages_per_fmr * sizeof (u64);
-
-               for (i = 0; i < params->pool_size; ++i) {
-                       fmr = kmalloc(bytes_per_fmr, GFP_KERNEL);
-                       if (!fmr)
-                               goto out_fail;
-
-                       fmr->pool             = pool;
-                       fmr->remap_count      = 0;
-                       fmr->ref_count        = 0;
-                       INIT_HLIST_NODE(&fmr->cache_node);
-
-                       fmr->fmr = ib_alloc_fmr(pd, params->access, &fmr_attr);
-                       if (IS_ERR(fmr->fmr)) {
-                               pr_warn(PFX "fmr_create failed for FMR %d\n",
-                                       i);
-                               kfree(fmr);
-                               goto out_fail;
-                       }
-
-                       list_add_tail(&fmr->list, &pool->free_list);
-                       ++pool->pool_size;
-               }
-       }
-
-       return pool;
-
- out_free_pool:
-       kfree(pool->cache_bucket);
-       kfree(pool);
-
-       return ERR_PTR(ret);
-
- out_fail:
-       ib_destroy_fmr_pool(pool);
-
-       return ERR_PTR(-ENOMEM);
-}
-EXPORT_SYMBOL(ib_create_fmr_pool);
-
-/**
- * ib_destroy_fmr_pool - Free FMR pool
- * @pool:FMR pool to free
- *
- * Destroy an FMR pool and free all associated resources.
- */
-void ib_destroy_fmr_pool(struct ib_fmr_pool *pool)
-{
-       struct ib_pool_fmr *fmr;
-       struct ib_pool_fmr *tmp;
-       LIST_HEAD(fmr_list);
-       int                 i;
-
-       kthread_destroy_worker(pool->worker);
-       ib_fmr_batch_release(pool);
-
-       i = 0;
-       list_for_each_entry_safe(fmr, tmp, &pool->free_list, list) {
-               if (fmr->remap_count) {
-                       INIT_LIST_HEAD(&fmr_list);
-                       list_add_tail(&fmr->fmr->list, &fmr_list);
-                       ib_unmap_fmr(&fmr_list);
-               }
-               ib_dealloc_fmr(fmr->fmr);
-               list_del(&fmr->list);
-               kfree(fmr);
-               ++i;
-       }
-
-       if (i < pool->pool_size)
-               pr_warn(PFX "pool still has %d regions registered\n",
-                       pool->pool_size - i);
-
-       kfree(pool->cache_bucket);
-       kfree(pool);
-}
-EXPORT_SYMBOL(ib_destroy_fmr_pool);
-
-/**
- * ib_flush_fmr_pool - Invalidate all unmapped FMRs
- * @pool:FMR pool to flush
- *
- * Ensure that all unmapped FMRs are fully invalidated.
- */
-int ib_flush_fmr_pool(struct ib_fmr_pool *pool)
-{
-       int serial;
-       struct ib_pool_fmr *fmr, *next;
-
-       /*
-        * The free_list holds FMRs that may have been used
-        * but have not been remapped enough times to be dirty.
-        * Put them on the dirty list now so that the cleanup
-        * thread will reap them too.
-        */
-       spin_lock_irq(&pool->pool_lock);
-       list_for_each_entry_safe(fmr, next, &pool->free_list, list) {
-               if (fmr->remap_count > 0)
-                       list_move(&fmr->list, &pool->dirty_list);
-       }
-       spin_unlock_irq(&pool->pool_lock);
-
-       serial = atomic_inc_return(&pool->req_ser);
-       kthread_queue_work(pool->worker, &pool->work);
-
-       if (wait_event_interruptible(pool->force_wait,
-                                    atomic_read(&pool->flush_ser) - serial >= 0))
-               return -EINTR;
-
-       return 0;
-}
-EXPORT_SYMBOL(ib_flush_fmr_pool);
-
-/**
- * ib_fmr_pool_map_phys - Map an FMR from an FMR pool.
- * @pool_handle: FMR pool to allocate FMR from
- * @page_list: List of pages to map
- * @list_len: Number of pages in @page_list
- * @io_virtual_address: I/O virtual address for new FMR
- */
-struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
-                                        u64                *page_list,
-                                        int                 list_len,
-                                        u64                 io_virtual_address)
-{
-       struct ib_fmr_pool *pool = pool_handle;
-       struct ib_pool_fmr *fmr;
-       unsigned long       flags;
-       int                 result;
-
-       if (list_len < 1 || list_len > pool->max_pages)
-               return ERR_PTR(-EINVAL);
-
-       spin_lock_irqsave(&pool->pool_lock, flags);
-       fmr = ib_fmr_cache_lookup(pool,
-                                 page_list,
-                                 list_len,
-                                 io_virtual_address);
-       if (fmr) {
-               /* found in cache */
-               ++fmr->ref_count;
-               if (fmr->ref_count == 1) {
-                       list_del(&fmr->list);
-               }
-
-               spin_unlock_irqrestore(&pool->pool_lock, flags);
-
-               return fmr;
-       }
-
-       if (list_empty(&pool->free_list)) {
-               spin_unlock_irqrestore(&pool->pool_lock, flags);
-               return ERR_PTR(-EAGAIN);
-       }
-
-       fmr = list_entry(pool->free_list.next, struct ib_pool_fmr, list);
-       list_del(&fmr->list);
-       hlist_del_init(&fmr->cache_node);
-       spin_unlock_irqrestore(&pool->pool_lock, flags);
-
-       result = ib_map_phys_fmr(fmr->fmr, page_list, list_len,
-                                io_virtual_address);
-
-       if (result) {
-               spin_lock_irqsave(&pool->pool_lock, flags);
-               list_add(&fmr->list, &pool->free_list);
-               spin_unlock_irqrestore(&pool->pool_lock, flags);
-
-               pr_warn(PFX "fmr_map returns %d\n", result);
-
-               return ERR_PTR(result);
-       }
-
-       ++fmr->remap_count;
-       fmr->ref_count = 1;
-
-       if (pool->cache_bucket) {
-               fmr->io_virtual_address = io_virtual_address;
-               fmr->page_list_len      = list_len;
-               memcpy(fmr->page_list, page_list, list_len * sizeof(*page_list));
-
-               spin_lock_irqsave(&pool->pool_lock, flags);
-               hlist_add_head(&fmr->cache_node,
-                              pool->cache_bucket + ib_fmr_hash(fmr->page_list[0]));
-               spin_unlock_irqrestore(&pool->pool_lock, flags);
-       }
-
-       return fmr;
-}
-EXPORT_SYMBOL(ib_fmr_pool_map_phys);
-
-/**
- * ib_fmr_pool_unmap - Unmap FMR
- * @fmr:FMR to unmap
- *
- * Unmap an FMR.  The FMR mapping may remain valid until the FMR is
- * reused (or until ib_flush_fmr_pool() is called).
- */
-void ib_fmr_pool_unmap(struct ib_pool_fmr *fmr)
-{
-       struct ib_fmr_pool *pool;
-       unsigned long flags;
-
-       pool = fmr->pool;
-
-       spin_lock_irqsave(&pool->pool_lock, flags);
-
-       --fmr->ref_count;
-       if (!fmr->ref_count) {
-               if (fmr->remap_count < pool->max_remaps) {
-                       list_add_tail(&fmr->list, &pool->free_list);
-               } else {
-                       list_add_tail(&fmr->list, &pool->dirty_list);
-                       if (++pool->dirty_len >= pool->dirty_watermark) {
-                               atomic_inc(&pool->req_ser);
-                               kthread_queue_work(pool->worker, &pool->work);
-                       }
-               }
-       }
-
-       spin_unlock_irqrestore(&pool->pool_lock, flags);
-}
-EXPORT_SYMBOL(ib_fmr_pool_unmap);
diff --git a/include/rdma/ib_fmr_pool.h b/include/rdma/ib_fmr_pool.h
deleted file mode 100644 (file)
index 2fd9bfb..0000000
+++ /dev/null
@@ -1,93 +0,0 @@
-/*
- * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
- * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
- *
- * This software is available to you under a choice of one of two
- * licenses.  You may choose to be licensed under the terms of the GNU
- * General Public License (GPL) Version 2, available from the file
- * COPYING in the main directory of this source tree, or the
- * OpenIB.org BSD license below:
- *
- *     Redistribution and use in source and binary forms, with or
- *     without modification, are permitted provided that the following
- *     conditions are met:
- *
- *      - Redistributions of source code must retain the above
- *        copyright notice, this list of conditions and the following
- *        disclaimer.
- *
- *      - Redistributions in binary form must reproduce the above
- *        copyright notice, this list of conditions and the following
- *        disclaimer in the documentation and/or other materials
- *        provided with the distribution.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
- * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
- * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
- * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
- * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
- * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-
-#if !defined(IB_FMR_POOL_H)
-#define IB_FMR_POOL_H
-
-#include <rdma/ib_verbs.h>
-
-struct ib_fmr_pool;
-
-/**
- * struct ib_fmr_pool_param - Parameters for creating FMR pool
- * @max_pages_per_fmr:Maximum number of pages per map request.
- * @page_shift: Log2 of sizeof "pages" mapped by this fmr
- * @access:Access flags for FMRs in pool.
- * @pool_size:Number of FMRs to allocate for pool.
- * @dirty_watermark:Flush is triggered when @dirty_watermark dirty
- *     FMRs are present.
- * @flush_function:Callback called when unmapped FMRs are flushed and
- *     more FMRs are possibly available for mapping
- * @flush_arg:Context passed to user's flush function.
- * @cache:If set, FMRs may be reused after unmapping for identical map
- *     requests.
- */
-struct ib_fmr_pool_param {
-       int                     max_pages_per_fmr;
-       int                     page_shift;
-       enum ib_access_flags    access;
-       int                     pool_size;
-       int                     dirty_watermark;
-       void                  (*flush_function)(struct ib_fmr_pool *pool,
-                                               void               *arg);
-       void                   *flush_arg;
-       unsigned                cache:1;
-};
-
-struct ib_pool_fmr {
-       struct ib_fmr      *fmr;
-       struct ib_fmr_pool *pool;
-       struct list_head    list;
-       struct hlist_node   cache_node;
-       int                 ref_count;
-       int                 remap_count;
-       u64                 io_virtual_address;
-       int                 page_list_len;
-       u64                 page_list[];
-};
-
-struct ib_fmr_pool *ib_create_fmr_pool(struct ib_pd             *pd,
-                                      struct ib_fmr_pool_param *params);
-
-void ib_destroy_fmr_pool(struct ib_fmr_pool *pool);
-
-int ib_flush_fmr_pool(struct ib_fmr_pool *pool);
-
-struct ib_pool_fmr *ib_fmr_pool_map_phys(struct ib_fmr_pool *pool_handle,
-                                        u64                *page_list,
-                                        int                 list_len,
-                                        u64                 io_virtual_address);
-
-void ib_fmr_pool_unmap(struct ib_pool_fmr *fmr);
-
-#endif /* IB_FMR_POOL_H */