misc: mic: SCIF memory registration and unregistration
authorSudeep Dutt <sudeep.dutt@intel.com>
Wed, 30 Sep 2015 01:15:29 +0000 (18:15 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 4 Oct 2015 11:54:54 +0000 (12:54 +0100)
This patch implements the SCIF APIs required to pin and unpin
pages. SCIF registration locks down the pages. It then sends a remote
window allocation request to the peer. Once the peer has allocated
memory, the local SCIF endpoint copies the pinned page information to
the peer and notifies the peer once the copy has complete. The peer
upon receipt of the registration notification adds the new remote
window to its list. At this point the window page information is
available on both self and remote nodes so that they can start
performing SCIF DMAs, CPU copies and fences. The unregistration API
tears down the registration at both self and remote nodes.

Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/misc/mic/scif/scif_rma.c [new file with mode: 0644]

diff --git a/drivers/misc/mic/scif/scif_rma.c b/drivers/misc/mic/scif/scif_rma.c
new file mode 100644 (file)
index 0000000..e288996
--- /dev/null
@@ -0,0 +1,1770 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include <linux/dma_remapping.h>
+#include <linux/pagemap.h>
+#include "scif_main.h"
+#include "scif_map.h"
+
+/* Used to skip ulimit checks for registrations with SCIF_MAP_KERNEL flag */
+#define SCIF_MAP_ULIMIT 0x40
+
+bool scif_ulimit_check = 1;
+
+/**
+ * scif_rma_ep_init:
+ * @ep: end point
+ *
+ * Initialize RMA per EP data structures.
+ */
+void scif_rma_ep_init(struct scif_endpt *ep)
+{
+       struct scif_endpt_rma_info *rma = &ep->rma_info;
+
+       mutex_init(&rma->rma_lock);
+       init_iova_domain(&rma->iovad, PAGE_SIZE, SCIF_IOVA_START_PFN,
+                        SCIF_DMA_64BIT_PFN);
+       spin_lock_init(&rma->tc_lock);
+       mutex_init(&rma->mmn_lock);
+       INIT_LIST_HEAD(&rma->reg_list);
+       INIT_LIST_HEAD(&rma->remote_reg_list);
+       atomic_set(&rma->tw_refcount, 0);
+       atomic_set(&rma->tcw_refcount, 0);
+       atomic_set(&rma->tcw_total_pages, 0);
+       atomic_set(&rma->fence_refcount, 0);
+
+       rma->async_list_del = 0;
+       rma->dma_chan = NULL;
+       INIT_LIST_HEAD(&rma->mmn_list);
+       INIT_LIST_HEAD(&rma->vma_list);
+       init_waitqueue_head(&rma->markwq);
+}
+
+/**
+ * scif_rma_ep_can_uninit:
+ * @ep: end point
+ *
+ * Returns 1 if an endpoint can be uninitialized and 0 otherwise.
+ */
+int scif_rma_ep_can_uninit(struct scif_endpt *ep)
+{
+       int ret = 0;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       /* Destroy RMA Info only if both lists are empty */
+       if (list_empty(&ep->rma_info.reg_list) &&
+           list_empty(&ep->rma_info.remote_reg_list) &&
+           list_empty(&ep->rma_info.mmn_list) &&
+           !atomic_read(&ep->rma_info.tw_refcount) &&
+           !atomic_read(&ep->rma_info.tcw_refcount) &&
+           !atomic_read(&ep->rma_info.fence_refcount))
+               ret = 1;
+       mutex_unlock(&ep->rma_info.rma_lock);
+       return ret;
+}
+
+/**
+ * scif_create_pinned_pages:
+ * @nr_pages: number of pages in window
+ * @prot: read/write protection
+ *
+ * Allocate and prepare a set of pinned pages.
+ */
+static struct scif_pinned_pages *
+scif_create_pinned_pages(int nr_pages, int prot)
+{
+       struct scif_pinned_pages *pin;
+
+       might_sleep();
+       pin = scif_zalloc(sizeof(*pin));
+       if (!pin)
+               goto error;
+
+       pin->pages = scif_zalloc(nr_pages * sizeof(*pin->pages));
+       if (!pin->pages)
+               goto error_free_pinned_pages;
+
+       pin->prot = prot;
+       pin->magic = SCIFEP_MAGIC;
+       return pin;
+
+error_free_pinned_pages:
+       scif_free(pin, sizeof(*pin));
+error:
+       return NULL;
+}
+
+/**
+ * scif_destroy_pinned_pages:
+ * @pin: A set of pinned pages.
+ *
+ * Deallocate resources for pinned pages.
+ */
+static int scif_destroy_pinned_pages(struct scif_pinned_pages *pin)
+{
+       int j;
+       int writeable = pin->prot & SCIF_PROT_WRITE;
+       int kernel = SCIF_MAP_KERNEL & pin->map_flags;
+
+       for (j = 0; j < pin->nr_pages; j++) {
+               if (pin->pages[j] && !kernel) {
+                       if (writeable)
+                               SetPageDirty(pin->pages[j]);
+                       put_page(pin->pages[j]);
+               }
+       }
+
+       scif_free(pin->pages,
+                 pin->nr_pages * sizeof(*pin->pages));
+       scif_free(pin, sizeof(*pin));
+       return 0;
+}
+
+/*
+ * scif_create_window:
+ * @ep: end point
+ * @nr_pages: number of pages
+ * @offset: registration offset
+ * @temp: true if a temporary window is being created
+ *
+ * Allocate and prepare a self registration window.
+ */
+struct scif_window *scif_create_window(struct scif_endpt *ep, int nr_pages,
+                                      s64 offset, bool temp)
+{
+       struct scif_window *window;
+
+       might_sleep();
+       window = scif_zalloc(sizeof(*window));
+       if (!window)
+               goto error;
+
+       window->dma_addr = scif_zalloc(nr_pages * sizeof(*window->dma_addr));
+       if (!window->dma_addr)
+               goto error_free_window;
+
+       window->num_pages = scif_zalloc(nr_pages * sizeof(*window->num_pages));
+       if (!window->num_pages)
+               goto error_free_window;
+
+       window->offset = offset;
+       window->ep = (u64)ep;
+       window->magic = SCIFEP_MAGIC;
+       window->reg_state = OP_IDLE;
+       init_waitqueue_head(&window->regwq);
+       window->unreg_state = OP_IDLE;
+       init_waitqueue_head(&window->unregwq);
+       INIT_LIST_HEAD(&window->list);
+       window->type = SCIF_WINDOW_SELF;
+       window->temp = temp;
+       return window;
+
+error_free_window:
+       scif_free(window->dma_addr,
+                 nr_pages * sizeof(*window->dma_addr));
+       scif_free(window, sizeof(*window));
+error:
+       return NULL;
+}
+
+/**
+ * scif_destroy_incomplete_window:
+ * @ep: end point
+ * @window: registration window
+ *
+ * Deallocate resources for self window.
+ */
+static void scif_destroy_incomplete_window(struct scif_endpt *ep,
+                                          struct scif_window *window)
+{
+       int err;
+       int nr_pages = window->nr_pages;
+       struct scif_allocmsg *alloc = &window->alloc_handle;
+       struct scifmsg msg;
+
+retry:
+       /* Wait for a SCIF_ALLOC_GNT/REJ message */
+       err = wait_event_timeout(alloc->allocwq,
+                                alloc->state != OP_IN_PROGRESS,
+                                SCIF_NODE_ALIVE_TIMEOUT);
+       if (!err && scifdev_alive(ep))
+               goto retry;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       if (alloc->state == OP_COMPLETED) {
+               msg.uop = SCIF_FREE_VIRT;
+               msg.src = ep->port;
+               msg.payload[0] = ep->remote_ep;
+               msg.payload[1] = window->alloc_handle.vaddr;
+               msg.payload[2] = (u64)window;
+               msg.payload[3] = SCIF_REGISTER;
+               _scif_nodeqp_send(ep->remote_dev, &msg);
+       }
+       mutex_unlock(&ep->rma_info.rma_lock);
+
+       scif_free_window_offset(ep, window, window->offset);
+       scif_free(window->dma_addr, nr_pages * sizeof(*window->dma_addr));
+       scif_free(window->num_pages, nr_pages * sizeof(*window->num_pages));
+       scif_free(window, sizeof(*window));
+}
+
+/**
+ * scif_unmap_window:
+ * @remote_dev: SCIF remote device
+ * @window: registration window
+ *
+ * Delete any DMA mappings created for a registered self window
+ */
+void scif_unmap_window(struct scif_dev *remote_dev, struct scif_window *window)
+{
+       int j;
+
+       if (scif_is_iommu_enabled() && !scifdev_self(remote_dev)) {
+               if (window->st) {
+                       dma_unmap_sg(&remote_dev->sdev->dev,
+                                    window->st->sgl, window->st->nents,
+                                    DMA_BIDIRECTIONAL);
+                       sg_free_table(window->st);
+                       kfree(window->st);
+                       window->st = NULL;
+               }
+       } else {
+               for (j = 0; j < window->nr_contig_chunks; j++) {
+                       if (window->dma_addr[j]) {
+                               scif_unmap_single(window->dma_addr[j],
+                                                 remote_dev,
+                                                 window->num_pages[j] <<
+                                                 PAGE_SHIFT);
+                               window->dma_addr[j] = 0x0;
+                       }
+               }
+       }
+}
+
+static inline struct mm_struct *__scif_acquire_mm(void)
+{
+       if (scif_ulimit_check)
+               return get_task_mm(current);
+       return NULL;
+}
+
+static inline void __scif_release_mm(struct mm_struct *mm)
+{
+       if (mm)
+               mmput(mm);
+}
+
+static inline int
+__scif_dec_pinned_vm_lock(struct mm_struct *mm,
+                         int nr_pages, bool try_lock)
+{
+       if (!mm || !nr_pages || !scif_ulimit_check)
+               return 0;
+       if (try_lock) {
+               if (!down_write_trylock(&mm->mmap_sem)) {
+                       dev_err(scif_info.mdev.this_device,
+                               "%s %d err\n", __func__, __LINE__);
+                       return -1;
+               }
+       } else {
+               down_write(&mm->mmap_sem);
+       }
+       mm->pinned_vm -= nr_pages;
+       up_write(&mm->mmap_sem);
+       return 0;
+}
+
+static inline int __scif_check_inc_pinned_vm(struct mm_struct *mm,
+                                            int nr_pages)
+{
+       unsigned long locked, lock_limit;
+
+       if (!mm || !nr_pages || !scif_ulimit_check)
+               return 0;
+
+       locked = nr_pages;
+       locked += mm->pinned_vm;
+       lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+       if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
+               dev_err(scif_info.mdev.this_device,
+                       "locked(%lu) > lock_limit(%lu)\n",
+                       locked, lock_limit);
+               return -ENOMEM;
+       }
+       mm->pinned_vm = locked;
+       return 0;
+}
+
+/**
+ * scif_destroy_window:
+ * @ep: end point
+ * @window: registration window
+ *
+ * Deallocate resources for self window.
+ */
+int scif_destroy_window(struct scif_endpt *ep, struct scif_window *window)
+{
+       int j;
+       struct scif_pinned_pages *pinned_pages = window->pinned_pages;
+       int nr_pages = window->nr_pages;
+
+       might_sleep();
+       if (!window->temp && window->mm) {
+               __scif_dec_pinned_vm_lock(window->mm, window->nr_pages, 0);
+               __scif_release_mm(window->mm);
+               window->mm = NULL;
+       }
+
+       scif_free_window_offset(ep, window, window->offset);
+       scif_unmap_window(ep->remote_dev, window);
+       /*
+        * Decrement references for this set of pinned pages from
+        * this window.
+        */
+       j = atomic_sub_return(1, &pinned_pages->ref_count);
+       if (j < 0)
+               dev_err(scif_info.mdev.this_device,
+                       "%s %d incorrect ref count %d\n",
+                       __func__, __LINE__, j);
+       /*
+        * If the ref count for pinned_pages is zero then someone
+        * has already called scif_unpin_pages() for it and we should
+        * destroy the page cache.
+        */
+       if (!j)
+               scif_destroy_pinned_pages(window->pinned_pages);
+       scif_free(window->dma_addr, nr_pages * sizeof(*window->dma_addr));
+       scif_free(window->num_pages, nr_pages * sizeof(*window->num_pages));
+       window->magic = 0;
+       scif_free(window, sizeof(*window));
+       return 0;
+}
+
+/**
+ * scif_create_remote_lookup:
+ * @remote_dev: SCIF remote device
+ * @window: remote window
+ *
+ * Allocate and prepare lookup entries for the remote
+ * end to copy over the physical addresses.
+ * Returns 0 on success and appropriate errno on failure.
+ */
+static int scif_create_remote_lookup(struct scif_dev *remote_dev,
+                                    struct scif_window *window)
+{
+       int i, j, err = 0;
+       int nr_pages = window->nr_pages;
+       bool vmalloc_dma_phys, vmalloc_num_pages;
+
+       might_sleep();
+       /* Map window */
+       err = scif_map_single(&window->mapped_offset,
+                             window, remote_dev, sizeof(*window));
+       if (err)
+               goto error_window;
+
+       /* Compute the number of lookup entries. 21 == 2MB Shift */
+       window->nr_lookup = ALIGN(nr_pages * PAGE_SIZE,
+                                       ((2) * 1024 * 1024)) >> 21;
+
+       window->dma_addr_lookup.lookup =
+               scif_alloc_coherent(&window->dma_addr_lookup.offset,
+                                   remote_dev, window->nr_lookup *
+                                   sizeof(*window->dma_addr_lookup.lookup),
+                                   GFP_KERNEL | __GFP_ZERO);
+       if (!window->dma_addr_lookup.lookup)
+               goto error_window;
+
+       window->num_pages_lookup.lookup =
+               scif_alloc_coherent(&window->num_pages_lookup.offset,
+                                   remote_dev, window->nr_lookup *
+                                   sizeof(*window->num_pages_lookup.lookup),
+                                   GFP_KERNEL | __GFP_ZERO);
+       if (!window->num_pages_lookup.lookup)
+               goto error_window;
+
+       vmalloc_dma_phys = is_vmalloc_addr(&window->dma_addr[0]);
+       vmalloc_num_pages = is_vmalloc_addr(&window->num_pages[0]);
+
+       /* Now map each of the pages containing physical addresses */
+       for (i = 0, j = 0; i < nr_pages; i += SCIF_NR_ADDR_IN_PAGE, j++) {
+               err = scif_map_page(&window->dma_addr_lookup.lookup[j],
+                                   vmalloc_dma_phys ?
+                                   vmalloc_to_page(&window->dma_addr[i]) :
+                                   virt_to_page(&window->dma_addr[i]),
+                                   remote_dev);
+               if (err)
+                       goto error_window;
+               err = scif_map_page(&window->num_pages_lookup.lookup[j],
+                                   vmalloc_dma_phys ?
+                                   vmalloc_to_page(&window->num_pages[i]) :
+                                   virt_to_page(&window->num_pages[i]),
+                                   remote_dev);
+               if (err)
+                       goto error_window;
+       }
+       return 0;
+error_window:
+       return err;
+}
+
+/**
+ * scif_destroy_remote_lookup:
+ * @remote_dev: SCIF remote device
+ * @window: remote window
+ *
+ * Destroy lookup entries used for the remote
+ * end to copy over the physical addresses.
+ */
+static void scif_destroy_remote_lookup(struct scif_dev *remote_dev,
+                                      struct scif_window *window)
+{
+       int i, j;
+
+       if (window->nr_lookup) {
+               struct scif_rma_lookup *lup = &window->dma_addr_lookup;
+               struct scif_rma_lookup *npup = &window->num_pages_lookup;
+
+               for (i = 0, j = 0; i < window->nr_pages;
+                       i += SCIF_NR_ADDR_IN_PAGE, j++) {
+                       if (lup->lookup && lup->lookup[j])
+                               scif_unmap_single(lup->lookup[j],
+                                                 remote_dev,
+                                                 PAGE_SIZE);
+                       if (npup->lookup && npup->lookup[j])
+                               scif_unmap_single(npup->lookup[j],
+                                                 remote_dev,
+                                                 PAGE_SIZE);
+               }
+               if (lup->lookup)
+                       scif_free_coherent(lup->lookup, lup->offset,
+                                          remote_dev, window->nr_lookup *
+                                          sizeof(*lup->lookup));
+               if (npup->lookup)
+                       scif_free_coherent(npup->lookup, npup->offset,
+                                          remote_dev, window->nr_lookup *
+                                          sizeof(*npup->lookup));
+               if (window->mapped_offset)
+                       scif_unmap_single(window->mapped_offset,
+                                         remote_dev, sizeof(*window));
+               window->nr_lookup = 0;
+       }
+}
+
+/**
+ * scif_create_remote_window:
+ * @ep: end point
+ * @nr_pages: number of pages in window
+ *
+ * Allocate and prepare a remote registration window.
+ */
+static struct scif_window *
+scif_create_remote_window(struct scif_dev *scifdev, int nr_pages)
+{
+       struct scif_window *window;
+
+       might_sleep();
+       window = scif_zalloc(sizeof(*window));
+       if (!window)
+               goto error_ret;
+
+       window->magic = SCIFEP_MAGIC;
+       window->nr_pages = nr_pages;
+
+       window->dma_addr = scif_zalloc(nr_pages * sizeof(*window->dma_addr));
+       if (!window->dma_addr)
+               goto error_window;
+
+       window->num_pages = scif_zalloc(nr_pages *
+                                       sizeof(*window->num_pages));
+       if (!window->num_pages)
+               goto error_window;
+
+       if (scif_create_remote_lookup(scifdev, window))
+               goto error_window;
+
+       window->type = SCIF_WINDOW_PEER;
+       window->unreg_state = OP_IDLE;
+       INIT_LIST_HEAD(&window->list);
+       return window;
+error_window:
+       scif_destroy_remote_window(window);
+error_ret:
+       return NULL;
+}
+
+/**
+ * scif_destroy_remote_window:
+ * @ep: end point
+ * @window: remote registration window
+ *
+ * Deallocate resources for remote window.
+ */
+void
+scif_destroy_remote_window(struct scif_window *window)
+{
+       scif_free(window->dma_addr, window->nr_pages *
+                 sizeof(*window->dma_addr));
+       scif_free(window->num_pages, window->nr_pages *
+                 sizeof(*window->num_pages));
+       window->magic = 0;
+       scif_free(window, sizeof(*window));
+}
+
+/**
+ * scif_iommu_map: create DMA mappings if the IOMMU is enabled
+ * @remote_dev: SCIF remote device
+ * @window: remote registration window
+ *
+ * Map the physical pages using dma_map_sg(..) and then detect the number
+ * of contiguous DMA mappings allocated
+ */
+static int scif_iommu_map(struct scif_dev *remote_dev,
+                         struct scif_window *window)
+{
+       struct scatterlist *sg;
+       int i, err;
+       scif_pinned_pages_t pin = window->pinned_pages;
+
+       window->st = kzalloc(sizeof(*window->st), GFP_KERNEL);
+       if (!window->st)
+               return -ENOMEM;
+
+       err = sg_alloc_table(window->st, window->nr_pages, GFP_KERNEL);
+       if (err)
+               return err;
+
+       for_each_sg(window->st->sgl, sg, window->st->nents, i)
+               sg_set_page(sg, pin->pages[i], PAGE_SIZE, 0x0);
+
+       err = dma_map_sg(&remote_dev->sdev->dev, window->st->sgl,
+                        window->st->nents, DMA_BIDIRECTIONAL);
+       if (!err)
+               return -ENOMEM;
+       /* Detect contiguous ranges of DMA mappings */
+       sg = window->st->sgl;
+       for (i = 0; sg; i++) {
+               dma_addr_t last_da;
+
+               window->dma_addr[i] = sg_dma_address(sg);
+               window->num_pages[i] = sg_dma_len(sg) >> PAGE_SHIFT;
+               last_da = sg_dma_address(sg) + sg_dma_len(sg);
+               while ((sg = sg_next(sg)) && sg_dma_address(sg) == last_da) {
+                       window->num_pages[i] +=
+                               (sg_dma_len(sg) >> PAGE_SHIFT);
+                       last_da = window->dma_addr[i] +
+                               sg_dma_len(sg);
+               }
+               window->nr_contig_chunks++;
+       }
+       return 0;
+}
+
+/**
+ * scif_map_window:
+ * @remote_dev: SCIF remote device
+ * @window: self registration window
+ *
+ * Map pages of a window into the aperture/PCI.
+ * Also determine addresses required for DMA.
+ */
+int
+scif_map_window(struct scif_dev *remote_dev, struct scif_window *window)
+{
+       int i, j, k, err = 0, nr_contig_pages;
+       scif_pinned_pages_t pin;
+       phys_addr_t phys_prev, phys_curr;
+
+       might_sleep();
+
+       pin = window->pinned_pages;
+
+       if (intel_iommu_enabled && !scifdev_self(remote_dev))
+               return scif_iommu_map(remote_dev, window);
+
+       for (i = 0, j = 0; i < window->nr_pages; i += nr_contig_pages, j++) {
+               phys_prev = page_to_phys(pin->pages[i]);
+               nr_contig_pages = 1;
+
+               /* Detect physically contiguous chunks */
+               for (k = i + 1; k < window->nr_pages; k++) {
+                       phys_curr = page_to_phys(pin->pages[k]);
+                       if (phys_curr != (phys_prev + PAGE_SIZE))
+                               break;
+                       phys_prev = phys_curr;
+                       nr_contig_pages++;
+               }
+               window->num_pages[j] = nr_contig_pages;
+               window->nr_contig_chunks++;
+               if (scif_is_mgmt_node()) {
+                       /*
+                        * Management node has to deal with SMPT on X100 and
+                        * hence the DMA mapping is required
+                        */
+                       err = scif_map_single(&window->dma_addr[j],
+                                             phys_to_virt(page_to_phys(
+                                                          pin->pages[i])),
+                                             remote_dev,
+                                             nr_contig_pages << PAGE_SHIFT);
+                       if (err)
+                               return err;
+               } else {
+                       window->dma_addr[j] = page_to_phys(pin->pages[i]);
+               }
+       }
+       return err;
+}
+
+/**
+ * scif_send_scif_unregister:
+ * @ep: end point
+ * @window: self registration window
+ *
+ * Send a SCIF_UNREGISTER message.
+ */
+static int scif_send_scif_unregister(struct scif_endpt *ep,
+                                    struct scif_window *window)
+{
+       struct scifmsg msg;
+
+       msg.uop = SCIF_UNREGISTER;
+       msg.src = ep->port;
+       msg.payload[0] = window->alloc_handle.vaddr;
+       msg.payload[1] = (u64)window;
+       return scif_nodeqp_send(ep->remote_dev, &msg);
+}
+
+/**
+ * scif_unregister_window:
+ * @window: self registration window
+ *
+ * Send an unregistration request and wait for a response.
+ */
+int scif_unregister_window(struct scif_window *window)
+{
+       int err = 0;
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+       bool send_msg = false;
+
+       might_sleep();
+       switch (window->unreg_state) {
+       case OP_IDLE:
+       {
+               window->unreg_state = OP_IN_PROGRESS;
+               send_msg = true;
+               /* fall through */
+       }
+       case OP_IN_PROGRESS:
+       {
+               scif_get_window(window, 1);
+               mutex_unlock(&ep->rma_info.rma_lock);
+               if (send_msg) {
+                       err = scif_send_scif_unregister(ep, window);
+                       if (err) {
+                               window->unreg_state = OP_COMPLETED;
+                               goto done;
+                       }
+               } else {
+                       /* Return ENXIO since unregistration is in progress */
+                       return -ENXIO;
+               }
+retry:
+               /* Wait for a SCIF_UNREGISTER_(N)ACK message */
+               err = wait_event_timeout(window->unregwq,
+                                        window->unreg_state != OP_IN_PROGRESS,
+                                        SCIF_NODE_ALIVE_TIMEOUT);
+               if (!err && scifdev_alive(ep))
+                       goto retry;
+               if (!err) {
+                       err = -ENODEV;
+                       window->unreg_state = OP_COMPLETED;
+                       dev_err(scif_info.mdev.this_device,
+                               "%s %d err %d\n", __func__, __LINE__, err);
+               }
+               if (err > 0)
+                       err = 0;
+done:
+               mutex_lock(&ep->rma_info.rma_lock);
+               scif_put_window(window, 1);
+               break;
+       }
+       case OP_FAILED:
+       {
+               if (!scifdev_alive(ep)) {
+                       err = -ENODEV;
+                       window->unreg_state = OP_COMPLETED;
+               }
+               break;
+       }
+       case OP_COMPLETED:
+               break;
+       default:
+               err = -ENODEV;
+       }
+
+       if (window->unreg_state == OP_COMPLETED && window->ref_count)
+               scif_put_window(window, window->nr_pages);
+
+       if (!window->ref_count) {
+               atomic_inc(&ep->rma_info.tw_refcount);
+               list_del_init(&window->list);
+               scif_free_window_offset(ep, window, window->offset);
+               mutex_unlock(&ep->rma_info.rma_lock);
+               if ((!!(window->pinned_pages->map_flags & SCIF_MAP_KERNEL)) &&
+                   scifdev_alive(ep)) {
+                       scif_drain_dma_intr(ep->remote_dev->sdev,
+                                           ep->rma_info.dma_chan);
+               } else {
+                       if (!__scif_dec_pinned_vm_lock(window->mm,
+                                                      window->nr_pages, 1)) {
+                               __scif_release_mm(window->mm);
+                               window->mm = NULL;
+                       }
+               }
+               scif_queue_for_cleanup(window, &scif_info.rma);
+               mutex_lock(&ep->rma_info.rma_lock);
+       }
+       return err;
+}
+
+/**
+ * scif_send_alloc_request:
+ * @ep: end point
+ * @window: self registration window
+ *
+ * Send a remote window allocation request
+ */
+static int scif_send_alloc_request(struct scif_endpt *ep,
+                                  struct scif_window *window)
+{
+       struct scifmsg msg;
+       struct scif_allocmsg *alloc = &window->alloc_handle;
+
+       /* Set up the Alloc Handle */
+       alloc->state = OP_IN_PROGRESS;
+       init_waitqueue_head(&alloc->allocwq);
+
+       /* Send out an allocation request */
+       msg.uop = SCIF_ALLOC_REQ;
+       msg.payload[1] = window->nr_pages;
+       msg.payload[2] = (u64)&window->alloc_handle;
+       return _scif_nodeqp_send(ep->remote_dev, &msg);
+}
+
+/**
+ * scif_prep_remote_window:
+ * @ep: end point
+ * @window: self registration window
+ *
+ * Send a remote window allocation request, wait for an allocation response,
+ * and prepares the remote window by copying over the page lists
+ */
+static int scif_prep_remote_window(struct scif_endpt *ep,
+                                  struct scif_window *window)
+{
+       struct scifmsg msg;
+       struct scif_window *remote_window;
+       struct scif_allocmsg *alloc = &window->alloc_handle;
+       dma_addr_t *dma_phys_lookup, *tmp, *num_pages_lookup, *tmp1;
+       int i = 0, j = 0;
+       int nr_contig_chunks, loop_nr_contig_chunks;
+       int remaining_nr_contig_chunks, nr_lookup;
+       int err, map_err;
+
+       map_err = scif_map_window(ep->remote_dev, window);
+       if (map_err)
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d map_err %d\n", __func__, __LINE__, map_err);
+       remaining_nr_contig_chunks = window->nr_contig_chunks;
+       nr_contig_chunks = window->nr_contig_chunks;
+retry:
+       /* Wait for a SCIF_ALLOC_GNT/REJ message */
+       err = wait_event_timeout(alloc->allocwq,
+                                alloc->state != OP_IN_PROGRESS,
+                                SCIF_NODE_ALIVE_TIMEOUT);
+       mutex_lock(&ep->rma_info.rma_lock);
+       /* Synchronize with the thread waking up allocwq */
+       mutex_unlock(&ep->rma_info.rma_lock);
+       if (!err && scifdev_alive(ep))
+               goto retry;
+
+       if (!err)
+               err = -ENODEV;
+
+       if (err > 0)
+               err = 0;
+       else
+               return err;
+
+       /* Bail out. The remote end rejected this request */
+       if (alloc->state == OP_FAILED)
+               return -ENOMEM;
+
+       if (map_err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, map_err);
+               msg.uop = SCIF_FREE_VIRT;
+               msg.src = ep->port;
+               msg.payload[0] = ep->remote_ep;
+               msg.payload[1] = window->alloc_handle.vaddr;
+               msg.payload[2] = (u64)window;
+               msg.payload[3] = SCIF_REGISTER;
+               spin_lock(&ep->lock);
+               if (ep->state == SCIFEP_CONNECTED)
+                       err = _scif_nodeqp_send(ep->remote_dev, &msg);
+               else
+                       err = -ENOTCONN;
+               spin_unlock(&ep->lock);
+               return err;
+       }
+
+       remote_window = scif_ioremap(alloc->phys_addr, sizeof(*window),
+                                    ep->remote_dev);
+
+       /* Compute the number of lookup entries. 21 == 2MB Shift */
+       nr_lookup = ALIGN(nr_contig_chunks, SCIF_NR_ADDR_IN_PAGE)
+                         >> ilog2(SCIF_NR_ADDR_IN_PAGE);
+
+       dma_phys_lookup =
+               scif_ioremap(remote_window->dma_addr_lookup.offset,
+                            nr_lookup *
+                            sizeof(*remote_window->dma_addr_lookup.lookup),
+                            ep->remote_dev);
+       num_pages_lookup =
+               scif_ioremap(remote_window->num_pages_lookup.offset,
+                            nr_lookup *
+                            sizeof(*remote_window->num_pages_lookup.lookup),
+                            ep->remote_dev);
+
+       while (remaining_nr_contig_chunks) {
+               loop_nr_contig_chunks = min_t(int, remaining_nr_contig_chunks,
+                                             (int)SCIF_NR_ADDR_IN_PAGE);
+               /* #1/2 - Copy  physical addresses over to the remote side */
+
+               /* #2/2 - Copy DMA addresses (addresses that are fed into the
+                * DMA engine) We transfer bus addresses which are then
+                * converted into a MIC physical address on the remote
+                * side if it is a MIC, if the remote node is a mgmt node we
+                * transfer the MIC physical address
+                */
+               tmp = scif_ioremap(dma_phys_lookup[j],
+                                  loop_nr_contig_chunks *
+                                  sizeof(*window->dma_addr),
+                                  ep->remote_dev);
+               tmp1 = scif_ioremap(num_pages_lookup[j],
+                                   loop_nr_contig_chunks *
+                                   sizeof(*window->num_pages),
+                                   ep->remote_dev);
+               if (scif_is_mgmt_node()) {
+                       memcpy_toio((void __force __iomem *)tmp,
+                                   &window->dma_addr[i], loop_nr_contig_chunks
+                                   * sizeof(*window->dma_addr));
+                       memcpy_toio((void __force __iomem *)tmp1,
+                                   &window->num_pages[i], loop_nr_contig_chunks
+                                   * sizeof(*window->num_pages));
+               } else {
+                       if (scifdev_is_p2p(ep->remote_dev)) {
+                               /*
+                                * add remote node's base address for this node
+                                * to convert it into a MIC address
+                                */
+                               int m;
+                               dma_addr_t dma_addr;
+
+                               for (m = 0; m < loop_nr_contig_chunks; m++) {
+                                       dma_addr = window->dma_addr[i + m] +
+                                               ep->remote_dev->base_addr;
+                                       writeq(dma_addr,
+                                              (void __force __iomem *)&tmp[m]);
+                               }
+                               memcpy_toio((void __force __iomem *)tmp1,
+                                           &window->num_pages[i],
+                                           loop_nr_contig_chunks
+                                           * sizeof(*window->num_pages));
+                       } else {
+                               /* Mgmt node or loopback - transfer DMA
+                                * addresses as is, this is the same as a
+                                * MIC physical address (we use the dma_addr
+                                * and not the phys_addr array since the
+                                * phys_addr is only setup if there is a mmap()
+                                * request from the mgmt node)
+                                */
+                               memcpy_toio((void __force __iomem *)tmp,
+                                           &window->dma_addr[i],
+                                           loop_nr_contig_chunks *
+                                           sizeof(*window->dma_addr));
+                               memcpy_toio((void __force __iomem *)tmp1,
+                                           &window->num_pages[i],
+                                           loop_nr_contig_chunks *
+                                           sizeof(*window->num_pages));
+                       }
+               }
+               remaining_nr_contig_chunks -= loop_nr_contig_chunks;
+               i += loop_nr_contig_chunks;
+               j++;
+               scif_iounmap(tmp, loop_nr_contig_chunks *
+                            sizeof(*window->dma_addr), ep->remote_dev);
+               scif_iounmap(tmp1, loop_nr_contig_chunks *
+                            sizeof(*window->num_pages), ep->remote_dev);
+       }
+
+       /* Prepare the remote window for the peer */
+       remote_window->peer_window = (u64)window;
+       remote_window->offset = window->offset;
+       remote_window->prot = window->prot;
+       remote_window->nr_contig_chunks = nr_contig_chunks;
+       remote_window->ep = ep->remote_ep;
+       scif_iounmap(num_pages_lookup,
+                    nr_lookup *
+                    sizeof(*remote_window->num_pages_lookup.lookup),
+                    ep->remote_dev);
+       scif_iounmap(dma_phys_lookup,
+                    nr_lookup *
+                    sizeof(*remote_window->dma_addr_lookup.lookup),
+                    ep->remote_dev);
+       scif_iounmap(remote_window, sizeof(*remote_window), ep->remote_dev);
+       window->peer_window = alloc->vaddr;
+       return err;
+}
+
+/**
+ * scif_send_scif_register:
+ * @ep: end point
+ * @window: self registration window
+ *
+ * Send a SCIF_REGISTER message if EP is connected and wait for a
+ * SCIF_REGISTER_(N)ACK message else send a SCIF_FREE_VIRT
+ * message so that the peer can free its remote window allocated earlier.
+ */
+static int scif_send_scif_register(struct scif_endpt *ep,
+                                  struct scif_window *window)
+{
+       int err = 0;
+       struct scifmsg msg;
+
+       msg.src = ep->port;
+       msg.payload[0] = ep->remote_ep;
+       msg.payload[1] = window->alloc_handle.vaddr;
+       msg.payload[2] = (u64)window;
+       spin_lock(&ep->lock);
+       if (ep->state == SCIFEP_CONNECTED) {
+               msg.uop = SCIF_REGISTER;
+               window->reg_state = OP_IN_PROGRESS;
+               err = _scif_nodeqp_send(ep->remote_dev, &msg);
+               spin_unlock(&ep->lock);
+               if (!err) {
+retry:
+                       /* Wait for a SCIF_REGISTER_(N)ACK message */
+                       err = wait_event_timeout(window->regwq,
+                                                window->reg_state !=
+                                                OP_IN_PROGRESS,
+                                                SCIF_NODE_ALIVE_TIMEOUT);
+                       if (!err && scifdev_alive(ep))
+                               goto retry;
+                       err = !err ? -ENODEV : 0;
+                       if (window->reg_state == OP_FAILED)
+                               err = -ENOTCONN;
+               }
+       } else {
+               msg.uop = SCIF_FREE_VIRT;
+               msg.payload[3] = SCIF_REGISTER;
+               err = _scif_nodeqp_send(ep->remote_dev, &msg);
+               spin_unlock(&ep->lock);
+               if (!err)
+                       err = -ENOTCONN;
+       }
+       return err;
+}
+
+/**
+ * scif_get_window_offset:
+ * @ep: end point descriptor
+ * @flags: flags
+ * @offset: offset hint
+ * @num_pages: number of pages
+ * @out_offset: computed offset returned by reference.
+ *
+ * Compute/Claim a new offset for this EP.
+ */
+int scif_get_window_offset(struct scif_endpt *ep, int flags, s64 offset,
+                          int num_pages, s64 *out_offset)
+{
+       s64 page_index;
+       struct iova *iova_ptr;
+       int err = 0;
+
+       if (flags & SCIF_MAP_FIXED) {
+               page_index = SCIF_IOVA_PFN(offset);
+               iova_ptr = reserve_iova(&ep->rma_info.iovad, page_index,
+                                       page_index + num_pages - 1);
+               if (!iova_ptr)
+                       err = -EADDRINUSE;
+       } else {
+               iova_ptr = alloc_iova(&ep->rma_info.iovad, num_pages,
+                                     SCIF_DMA_63BIT_PFN - 1, 0);
+               if (!iova_ptr)
+                       err = -ENOMEM;
+       }
+       if (!err)
+               *out_offset = (iova_ptr->pfn_lo) << PAGE_SHIFT;
+       return err;
+}
+
+/**
+ * scif_free_window_offset:
+ * @ep: end point descriptor
+ * @window: registration window
+ * @offset: Offset to be freed
+ *
+ * Free offset for this EP. The callee is supposed to grab
+ * the RMA mutex before calling this API.
+ */
+void scif_free_window_offset(struct scif_endpt *ep,
+                            struct scif_window *window, s64 offset)
+{
+       if ((window && !window->offset_freed) || !window) {
+               free_iova(&ep->rma_info.iovad, offset >> PAGE_SHIFT);
+               if (window)
+                       window->offset_freed = true;
+       }
+}
+
+/**
+ * scif_alloc_req: Respond to SCIF_ALLOC_REQ interrupt message
+ * @msg:        Interrupt message
+ *
+ * Remote side is requesting a memory allocation.
+ */
+void scif_alloc_req(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       int err;
+       struct scif_window *window = NULL;
+       int nr_pages = msg->payload[1];
+
+       window = scif_create_remote_window(scifdev, nr_pages);
+       if (!window) {
+               err = -ENOMEM;
+               goto error;
+       }
+
+       /* The peer's allocation request is granted */
+       msg->uop = SCIF_ALLOC_GNT;
+       msg->payload[0] = (u64)window;
+       msg->payload[1] = window->mapped_offset;
+       err = scif_nodeqp_send(scifdev, msg);
+       if (err)
+               scif_destroy_remote_window(window);
+       return;
+error:
+       /* The peer's allocation request is rejected */
+       dev_err(&scifdev->sdev->dev,
+               "%s %d error %d alloc_ptr %p nr_pages 0x%x\n",
+               __func__, __LINE__, err, window, nr_pages);
+       msg->uop = SCIF_ALLOC_REJ;
+       scif_nodeqp_send(scifdev, msg);
+}
+
+/**
+ * scif_alloc_gnt_rej: Respond to SCIF_ALLOC_GNT/REJ interrupt message
+ * @msg:        Interrupt message
+ *
+ * Remote side responded to a memory allocation.
+ */
+void scif_alloc_gnt_rej(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_allocmsg *handle = (struct scif_allocmsg *)msg->payload[2];
+       struct scif_window *window = container_of(handle, struct scif_window,
+                                                 alloc_handle);
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       handle->vaddr = msg->payload[0];
+       handle->phys_addr = msg->payload[1];
+       if (msg->uop == SCIF_ALLOC_GNT)
+               handle->state = OP_COMPLETED;
+       else
+               handle->state = OP_FAILED;
+       wake_up(&handle->allocwq);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+/**
+ * scif_free_virt: Respond to SCIF_FREE_VIRT interrupt message
+ * @msg:        Interrupt message
+ *
+ * Free up memory kmalloc'd earlier.
+ */
+void scif_free_virt(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_window *window = (struct scif_window *)msg->payload[1];
+
+       scif_destroy_remote_window(window);
+}
+
+static void
+scif_fixup_aper_base(struct scif_dev *dev, struct scif_window *window)
+{
+       int j;
+       struct scif_hw_dev *sdev = dev->sdev;
+       phys_addr_t apt_base = 0;
+
+       /*
+        * Add the aperture base if the DMA address is not card relative
+        * since the DMA addresses need to be an offset into the bar
+        */
+       if (!scifdev_self(dev) && window->type == SCIF_WINDOW_PEER &&
+           sdev->aper && !sdev->card_rel_da)
+               apt_base = sdev->aper->pa;
+       else
+               return;
+
+       for (j = 0; j < window->nr_contig_chunks; j++) {
+               if (window->num_pages[j])
+                       window->dma_addr[j] += apt_base;
+               else
+                       break;
+       }
+}
+
+/**
+ * scif_recv_reg: Respond to SCIF_REGISTER interrupt message
+ * @msg:        Interrupt message
+ *
+ * Update remote window list with a new registered window.
+ */
+void scif_recv_reg(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+       struct scif_window *window =
+               (struct scif_window *)msg->payload[1];
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       spin_lock(&ep->lock);
+       if (ep->state == SCIFEP_CONNECTED) {
+               msg->uop = SCIF_REGISTER_ACK;
+               scif_nodeqp_send(ep->remote_dev, msg);
+               scif_fixup_aper_base(ep->remote_dev, window);
+               /* No further failures expected. Insert new window */
+               scif_insert_window(window, &ep->rma_info.remote_reg_list);
+       } else {
+               msg->uop = SCIF_REGISTER_NACK;
+               scif_nodeqp_send(ep->remote_dev, msg);
+       }
+       spin_unlock(&ep->lock);
+       mutex_unlock(&ep->rma_info.rma_lock);
+       /* free up any lookup resources now that page lists are transferred */
+       scif_destroy_remote_lookup(ep->remote_dev, window);
+       /*
+        * We could not insert the window but we need to
+        * destroy the window.
+        */
+       if (msg->uop == SCIF_REGISTER_NACK)
+               scif_destroy_remote_window(window);
+}
+
+/**
+ * scif_recv_unreg: Respond to SCIF_UNREGISTER interrupt message
+ * @msg:        Interrupt message
+ *
+ * Remove window from remote registration list;
+ */
+void scif_recv_unreg(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_rma_req req;
+       struct scif_window *window = NULL;
+       struct scif_window *recv_window =
+               (struct scif_window *)msg->payload[0];
+       struct scif_endpt *ep;
+       int del_window = 0;
+
+       ep = (struct scif_endpt *)recv_window->ep;
+       req.out_window = &window;
+       req.offset = recv_window->offset;
+       req.prot = 0;
+       req.nr_bytes = recv_window->nr_pages << PAGE_SHIFT;
+       req.type = SCIF_WINDOW_FULL;
+       req.head = &ep->rma_info.remote_reg_list;
+       msg->payload[0] = ep->remote_ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       /* Does a valid window exist? */
+       if (scif_query_window(&req)) {
+               dev_err(&scifdev->sdev->dev,
+                       "%s %d -ENXIO\n", __func__, __LINE__);
+               msg->uop = SCIF_UNREGISTER_ACK;
+               goto error;
+       }
+       if (window) {
+               if (window->ref_count)
+                       scif_put_window(window, window->nr_pages);
+               else
+                       dev_err(&scifdev->sdev->dev,
+                               "%s %d ref count should be +ve\n",
+                               __func__, __LINE__);
+               window->unreg_state = OP_COMPLETED;
+               if (!window->ref_count) {
+                       msg->uop = SCIF_UNREGISTER_ACK;
+                       atomic_inc(&ep->rma_info.tw_refcount);
+                       ep->rma_info.async_list_del = 1;
+                       list_del_init(&window->list);
+                       del_window = 1;
+               } else {
+                       /* NACK! There are valid references to this window */
+                       msg->uop = SCIF_UNREGISTER_NACK;
+               }
+       } else {
+               /* The window did not make its way to the list at all. ACK */
+               msg->uop = SCIF_UNREGISTER_ACK;
+               scif_destroy_remote_window(recv_window);
+       }
+error:
+       mutex_unlock(&ep->rma_info.rma_lock);
+       if (del_window)
+               scif_drain_dma_intr(ep->remote_dev->sdev,
+                                   ep->rma_info.dma_chan);
+       scif_nodeqp_send(ep->remote_dev, msg);
+       if (del_window)
+               scif_queue_for_cleanup(window, &scif_info.rma);
+}
+
+/**
+ * scif_recv_reg_ack: Respond to SCIF_REGISTER_ACK interrupt message
+ * @msg:        Interrupt message
+ *
+ * Wake up the window waiting to complete registration.
+ */
+void scif_recv_reg_ack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_window *window =
+               (struct scif_window *)msg->payload[2];
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       window->reg_state = OP_COMPLETED;
+       wake_up(&window->regwq);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+/**
+ * scif_recv_reg_nack: Respond to SCIF_REGISTER_NACK interrupt message
+ * @msg:        Interrupt message
+ *
+ * Wake up the window waiting to inform it that registration
+ * cannot be completed.
+ */
+void scif_recv_reg_nack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_window *window =
+               (struct scif_window *)msg->payload[2];
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       window->reg_state = OP_FAILED;
+       wake_up(&window->regwq);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+/**
+ * scif_recv_unreg_ack: Respond to SCIF_UNREGISTER_ACK interrupt message
+ * @msg:        Interrupt message
+ *
+ * Wake up the window waiting to complete unregistration.
+ */
+void scif_recv_unreg_ack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_window *window =
+               (struct scif_window *)msg->payload[1];
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       window->unreg_state = OP_COMPLETED;
+       wake_up(&window->unregwq);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+/**
+ * scif_recv_unreg_nack: Respond to SCIF_UNREGISTER_NACK interrupt message
+ * @msg:        Interrupt message
+ *
+ * Wake up the window waiting to inform it that unregistration
+ * cannot be completed immediately.
+ */
+void scif_recv_unreg_nack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+       struct scif_window *window =
+               (struct scif_window *)msg->payload[1];
+       struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+
+       mutex_lock(&ep->rma_info.rma_lock);
+       window->unreg_state = OP_FAILED;
+       wake_up(&window->unregwq);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+int __scif_pin_pages(void *addr, size_t len, int *out_prot,
+                    int map_flags, scif_pinned_pages_t *pages)
+{
+       struct scif_pinned_pages *pinned_pages;
+       int nr_pages, err = 0, i;
+       bool vmalloc_addr = false;
+       bool try_upgrade = false;
+       int prot = *out_prot;
+       int ulimit = 0;
+       struct mm_struct *mm = NULL;
+
+       /* Unsupported flags */
+       if (map_flags & ~(SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT))
+               return -EINVAL;
+       ulimit = !!(map_flags & SCIF_MAP_ULIMIT);
+
+       /* Unsupported protection requested */
+       if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE))
+               return -EINVAL;
+
+       /* addr/len must be page aligned. len should be non zero */
+       if (!len ||
+           (ALIGN((u64)addr, PAGE_SIZE) != (u64)addr) ||
+           (ALIGN((u64)len, PAGE_SIZE) != (u64)len))
+               return -EINVAL;
+
+       might_sleep();
+
+       nr_pages = len >> PAGE_SHIFT;
+
+       /* Allocate a set of pinned pages */
+       pinned_pages = scif_create_pinned_pages(nr_pages, prot);
+       if (!pinned_pages)
+               return -ENOMEM;
+
+       if (map_flags & SCIF_MAP_KERNEL) {
+               if (is_vmalloc_addr(addr))
+                       vmalloc_addr = true;
+
+               for (i = 0; i < nr_pages; i++) {
+                       if (vmalloc_addr)
+                               pinned_pages->pages[i] =
+                                       vmalloc_to_page(addr + (i * PAGE_SIZE));
+                       else
+                               pinned_pages->pages[i] =
+                                       virt_to_page(addr + (i * PAGE_SIZE));
+               }
+               pinned_pages->nr_pages = nr_pages;
+               pinned_pages->map_flags = SCIF_MAP_KERNEL;
+       } else {
+               /*
+                * SCIF supports registration caching. If a registration has
+                * been requested with read only permissions, then we try
+                * to pin the pages with RW permissions so that a subsequent
+                * transfer with RW permission can hit the cache instead of
+                * invalidating it. If the upgrade fails with RW then we
+                * revert back to R permission and retry
+                */
+               if (prot == SCIF_PROT_READ)
+                       try_upgrade = true;
+               prot |= SCIF_PROT_WRITE;
+retry:
+               mm = current->mm;
+               down_write(&mm->mmap_sem);
+               if (ulimit) {
+                       err = __scif_check_inc_pinned_vm(mm, nr_pages);
+                       if (err) {
+                               up_write(&mm->mmap_sem);
+                               pinned_pages->nr_pages = 0;
+                               goto error_unmap;
+                       }
+               }
+
+               pinned_pages->nr_pages = get_user_pages(
+                               current,
+                               mm,
+                               (u64)addr,
+                               nr_pages,
+                               !!(prot & SCIF_PROT_WRITE),
+                               0,
+                               pinned_pages->pages,
+                               NULL);
+               up_write(&mm->mmap_sem);
+               if (nr_pages != pinned_pages->nr_pages) {
+                       if (try_upgrade) {
+                               if (ulimit)
+                                       __scif_dec_pinned_vm_lock(mm,
+                                                                 nr_pages, 0);
+                               /* Roll back any pinned pages */
+                               for (i = 0; i < pinned_pages->nr_pages; i++) {
+                                       if (pinned_pages->pages[i])
+                                               put_page(
+                                               pinned_pages->pages[i]);
+                               }
+                               prot &= ~SCIF_PROT_WRITE;
+                               try_upgrade = false;
+                               goto retry;
+                       }
+               }
+               pinned_pages->map_flags = 0;
+       }
+
+       if (pinned_pages->nr_pages < nr_pages) {
+               err = -EFAULT;
+               pinned_pages->nr_pages = nr_pages;
+               goto dec_pinned;
+       }
+
+       *out_prot = prot;
+       atomic_set(&pinned_pages->ref_count, 1);
+       *pages = pinned_pages;
+       return err;
+dec_pinned:
+       if (ulimit)
+               __scif_dec_pinned_vm_lock(mm, nr_pages, 0);
+       /* Something went wrong! Rollback */
+error_unmap:
+       pinned_pages->nr_pages = nr_pages;
+       scif_destroy_pinned_pages(pinned_pages);
+       *pages = NULL;
+       dev_dbg(scif_info.mdev.this_device,
+               "%s %d err %d len 0x%lx\n", __func__, __LINE__, err, len);
+       return err;
+}
+
+int scif_pin_pages(void *addr, size_t len, int prot,
+                  int map_flags, scif_pinned_pages_t *pages)
+{
+       return __scif_pin_pages(addr, len, &prot, map_flags, pages);
+}
+EXPORT_SYMBOL_GPL(scif_pin_pages);
+
+int scif_unpin_pages(scif_pinned_pages_t pinned_pages)
+{
+       int err = 0, ret;
+
+       if (!pinned_pages || SCIFEP_MAGIC != pinned_pages->magic)
+               return -EINVAL;
+
+       ret = atomic_sub_return(1, &pinned_pages->ref_count);
+       if (ret < 0) {
+               dev_err(scif_info.mdev.this_device,
+                       "%s %d scif_unpin_pages called without pinning? rc %d\n",
+                       __func__, __LINE__, ret);
+               return -EINVAL;
+       }
+       /*
+        * Destroy the window if the ref count for this set of pinned
+        * pages has dropped to zero. If it is positive then there is
+        * a valid registered window which is backed by these pages and
+        * it will be destroyed once all such windows are unregistered.
+        */
+       if (!ret)
+               err = scif_destroy_pinned_pages(pinned_pages);
+
+       return err;
+}
+EXPORT_SYMBOL_GPL(scif_unpin_pages);
+
+static inline void
+scif_insert_local_window(struct scif_window *window, struct scif_endpt *ep)
+{
+       mutex_lock(&ep->rma_info.rma_lock);
+       scif_insert_window(window, &ep->rma_info.reg_list);
+       mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+off_t scif_register_pinned_pages(scif_epd_t epd,
+                                scif_pinned_pages_t pinned_pages,
+                                off_t offset, int map_flags)
+{
+       struct scif_endpt *ep = (struct scif_endpt *)epd;
+       s64 computed_offset;
+       struct scif_window *window;
+       int err;
+       size_t len;
+       struct device *spdev;
+
+       /* Unsupported flags */
+       if (map_flags & ~SCIF_MAP_FIXED)
+               return -EINVAL;
+
+       len = pinned_pages->nr_pages << PAGE_SHIFT;
+
+       /*
+        * Offset is not page aligned/negative or offset+len
+        * wraps around with SCIF_MAP_FIXED.
+        */
+       if ((map_flags & SCIF_MAP_FIXED) &&
+           ((ALIGN(offset, PAGE_SIZE) != offset) ||
+           (offset < 0) ||
+           (offset + (off_t)len < offset)))
+               return -EINVAL;
+
+       might_sleep();
+
+       err = scif_verify_epd(ep);
+       if (err)
+               return err;
+       /*
+        * It is an error to pass pinned_pages to scif_register_pinned_pages()
+        * after calling scif_unpin_pages().
+        */
+       if (!atomic_add_unless(&pinned_pages->ref_count, 1, 0))
+               return -EINVAL;
+
+       /* Compute the offset for this registration */
+       err = scif_get_window_offset(ep, map_flags, offset,
+                                    len, &computed_offset);
+       if (err) {
+               atomic_sub(1, &pinned_pages->ref_count);
+               return err;
+       }
+
+       /* Allocate and prepare self registration window */
+       window = scif_create_window(ep, pinned_pages->nr_pages,
+                                   computed_offset, false);
+       if (!window) {
+               atomic_sub(1, &pinned_pages->ref_count);
+               scif_free_window_offset(ep, NULL, computed_offset);
+               return -ENOMEM;
+       }
+
+       window->pinned_pages = pinned_pages;
+       window->nr_pages = pinned_pages->nr_pages;
+       window->prot = pinned_pages->prot;
+
+       spdev = scif_get_peer_dev(ep->remote_dev);
+       if (IS_ERR(spdev)) {
+               err = PTR_ERR(spdev);
+               scif_destroy_window(ep, window);
+               return err;
+       }
+       err = scif_send_alloc_request(ep, window);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, err);
+               goto error_unmap;
+       }
+
+       /* Prepare the remote registration window */
+       err = scif_prep_remote_window(ep, window);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, err);
+               goto error_unmap;
+       }
+
+       /* Tell the peer about the new window */
+       err = scif_send_scif_register(ep, window);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, err);
+               goto error_unmap;
+       }
+
+       scif_put_peer_dev(spdev);
+       /* No further failures expected. Insert new window */
+       scif_insert_local_window(window, ep);
+       return computed_offset;
+error_unmap:
+       scif_destroy_window(ep, window);
+       scif_put_peer_dev(spdev);
+       dev_err(&ep->remote_dev->sdev->dev,
+               "%s %d err %d\n", __func__, __LINE__, err);
+       return err;
+}
+EXPORT_SYMBOL_GPL(scif_register_pinned_pages);
+
+off_t scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset,
+                   int prot, int map_flags)
+{
+       scif_pinned_pages_t pinned_pages;
+       off_t err;
+       struct scif_endpt *ep = (struct scif_endpt *)epd;
+       s64 computed_offset;
+       struct scif_window *window;
+       struct mm_struct *mm = NULL;
+       struct device *spdev;
+
+       dev_dbg(scif_info.mdev.this_device,
+               "SCIFAPI register: ep %p addr %p len 0x%lx offset 0x%lx prot 0x%x map_flags 0x%x\n",
+               epd, addr, len, offset, prot, map_flags);
+       /* Unsupported flags */
+       if (map_flags & ~(SCIF_MAP_FIXED | SCIF_MAP_KERNEL))
+               return -EINVAL;
+
+       /*
+        * Offset is not page aligned/negative or offset+len
+        * wraps around with SCIF_MAP_FIXED.
+        */
+       if ((map_flags & SCIF_MAP_FIXED) &&
+           ((ALIGN(offset, PAGE_SIZE) != offset) ||
+           (offset < 0) ||
+           (offset + (off_t)len < offset)))
+               return -EINVAL;
+
+       /* Unsupported protection requested */
+       if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE))
+               return -EINVAL;
+
+       /* addr/len must be page aligned. len should be non zero */
+       if (!len || (ALIGN((u64)addr, PAGE_SIZE) != (u64)addr) ||
+           (ALIGN(len, PAGE_SIZE) != len))
+               return -EINVAL;
+
+       might_sleep();
+
+       err = scif_verify_epd(ep);
+       if (err)
+               return err;
+
+       /* Compute the offset for this registration */
+       err = scif_get_window_offset(ep, map_flags, offset,
+                                    len >> PAGE_SHIFT, &computed_offset);
+       if (err)
+               return err;
+
+       spdev = scif_get_peer_dev(ep->remote_dev);
+       if (IS_ERR(spdev)) {
+               err = PTR_ERR(spdev);
+               scif_free_window_offset(ep, NULL, computed_offset);
+               return err;
+       }
+       /* Allocate and prepare self registration window */
+       window = scif_create_window(ep, len >> PAGE_SHIFT,
+                                   computed_offset, false);
+       if (!window) {
+               scif_free_window_offset(ep, NULL, computed_offset);
+               scif_put_peer_dev(spdev);
+               return -ENOMEM;
+       }
+
+       window->nr_pages = len >> PAGE_SHIFT;
+
+       err = scif_send_alloc_request(ep, window);
+       if (err) {
+               scif_destroy_incomplete_window(ep, window);
+               scif_put_peer_dev(spdev);
+               return err;
+       }
+
+       if (!(map_flags & SCIF_MAP_KERNEL)) {
+               mm = __scif_acquire_mm();
+               map_flags |= SCIF_MAP_ULIMIT;
+       }
+       /* Pin down the pages */
+       err = __scif_pin_pages(addr, len, &prot,
+                              map_flags & (SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT),
+                              &pinned_pages);
+       if (err) {
+               scif_destroy_incomplete_window(ep, window);
+               __scif_release_mm(mm);
+               goto error;
+       }
+
+       window->pinned_pages = pinned_pages;
+       window->prot = pinned_pages->prot;
+       window->mm = mm;
+
+       /* Prepare the remote registration window */
+       err = scif_prep_remote_window(ep, window);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %ld\n", __func__, __LINE__, err);
+               goto error_unmap;
+       }
+
+       /* Tell the peer about the new window */
+       err = scif_send_scif_register(ep, window);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %ld\n", __func__, __LINE__, err);
+               goto error_unmap;
+       }
+
+       scif_put_peer_dev(spdev);
+       /* No further failures expected. Insert new window */
+       scif_insert_local_window(window, ep);
+       dev_dbg(&ep->remote_dev->sdev->dev,
+               "SCIFAPI register: ep %p addr %p len 0x%lx computed_offset 0x%llx\n",
+               epd, addr, len, computed_offset);
+       return computed_offset;
+error_unmap:
+       scif_destroy_window(ep, window);
+error:
+       scif_put_peer_dev(spdev);
+       dev_err(&ep->remote_dev->sdev->dev,
+               "%s %d err %ld\n", __func__, __LINE__, err);
+       return err;
+}
+EXPORT_SYMBOL_GPL(scif_register);
+
+int
+scif_unregister(scif_epd_t epd, off_t offset, size_t len)
+{
+       struct scif_endpt *ep = (struct scif_endpt *)epd;
+       struct scif_window *window = NULL;
+       struct scif_rma_req req;
+       int nr_pages, err;
+       struct device *spdev;
+
+       dev_dbg(scif_info.mdev.this_device,
+               "SCIFAPI unregister: ep %p offset 0x%lx len 0x%lx\n",
+               ep, offset, len);
+       /* len must be page aligned. len should be non zero */
+       if (!len ||
+           (ALIGN((u64)len, PAGE_SIZE) != (u64)len))
+               return -EINVAL;
+
+       /* Offset is not page aligned or offset+len wraps around */
+       if ((ALIGN(offset, PAGE_SIZE) != offset) ||
+           (offset + (off_t)len < offset))
+               return -EINVAL;
+
+       err = scif_verify_epd(ep);
+       if (err)
+               return err;
+
+       might_sleep();
+       nr_pages = len >> PAGE_SHIFT;
+
+       req.out_window = &window;
+       req.offset = offset;
+       req.prot = 0;
+       req.nr_bytes = len;
+       req.type = SCIF_WINDOW_FULL;
+       req.head = &ep->rma_info.reg_list;
+
+       spdev = scif_get_peer_dev(ep->remote_dev);
+       if (IS_ERR(spdev)) {
+               err = PTR_ERR(spdev);
+               return err;
+       }
+       mutex_lock(&ep->rma_info.rma_lock);
+       /* Does a valid window exist? */
+       err = scif_query_window(&req);
+       if (err) {
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, err);
+               goto error;
+       }
+       /* Unregister all the windows in this range */
+       err = scif_rma_list_unregister(window, offset, nr_pages);
+       if (err)
+               dev_err(&ep->remote_dev->sdev->dev,
+                       "%s %d err %d\n", __func__, __LINE__, err);
+error:
+       mutex_unlock(&ep->rma_info.rma_lock);
+       scif_put_peer_dev(spdev);
+       return err;
+}
+EXPORT_SYMBOL_GPL(scif_unregister);