#include <rdma/ib_umem_odp.h>
#include "mlx5_ib.h"
+#include "cmd.h"
#define MAX_PREFETCH_LEN (4*1024*1024U)
* a pagefault. */
#define MMU_NOTIFIER_TIMEOUT 1000
+#define MLX5_IMR_MTT_BITS (30 - PAGE_SHIFT)
+#define MLX5_IMR_MTT_SHIFT (MLX5_IMR_MTT_BITS + PAGE_SHIFT)
+#define MLX5_IMR_MTT_ENTRIES BIT_ULL(MLX5_IMR_MTT_BITS)
+#define MLX5_IMR_MTT_SIZE BIT_ULL(MLX5_IMR_MTT_SHIFT)
+#define MLX5_IMR_MTT_MASK (~(MLX5_IMR_MTT_SIZE - 1))
+
+#define MLX5_KSM_PAGE_SHIFT MLX5_IMR_MTT_SHIFT
+
+static u64 mlx5_imr_ksm_entries;
+
+static int check_parent(struct ib_umem_odp *odp,
+ struct mlx5_ib_mr *parent)
+{
+ struct mlx5_ib_mr *mr = odp->private;
+
+ return mr && mr->parent == parent;
+}
+
+static struct ib_umem_odp *odp_next(struct ib_umem_odp *odp)
+{
+ struct mlx5_ib_mr *mr = odp->private, *parent = mr->parent;
+ struct ib_ucontext *ctx = odp->umem->context;
+ struct rb_node *rb;
+
+ down_read(&ctx->umem_rwsem);
+ while (1) {
+ rb = rb_next(&odp->interval_tree.rb);
+ if (!rb)
+ goto not_found;
+ odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
+ if (check_parent(odp, parent))
+ goto end;
+ }
+not_found:
+ odp = NULL;
+end:
+ up_read(&ctx->umem_rwsem);
+ return odp;
+}
+
+static struct ib_umem_odp *odp_lookup(struct ib_ucontext *ctx,
+ u64 start, u64 length,
+ struct mlx5_ib_mr *parent)
+{
+ struct ib_umem_odp *odp;
+ struct rb_node *rb;
+
+ down_read(&ctx->umem_rwsem);
+ odp = rbt_ib_umem_lookup(&ctx->umem_tree, start, length);
+ if (!odp)
+ goto end;
+
+ while (1) {
+ if (check_parent(odp, parent))
+ goto end;
+ rb = rb_next(&odp->interval_tree.rb);
+ if (!rb)
+ goto not_found;
+ odp = rb_entry(rb, struct ib_umem_odp, interval_tree.rb);
+ if (ib_umem_start(odp->umem) > start + length)
+ goto not_found;
+ }
+not_found:
+ odp = NULL;
+end:
+ up_read(&ctx->umem_rwsem);
+ return odp;
+}
+
+void mlx5_odp_populate_klm(struct mlx5_klm *pklm, size_t offset,
+ size_t nentries, struct mlx5_ib_mr *mr, int flags)
+{
+ struct ib_pd *pd = mr->ibmr.pd;
+ struct ib_ucontext *ctx = pd->uobject->context;
+ struct mlx5_ib_dev *dev = to_mdev(pd->device);
+ struct ib_umem_odp *odp;
+ unsigned long va;
+ int i;
+
+ if (flags & MLX5_IB_UPD_XLT_ZAP) {
+ for (i = 0; i < nentries; i++, pklm++) {
+ pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
+ pklm->key = cpu_to_be32(dev->null_mkey);
+ pklm->va = 0;
+ }
+ return;
+ }
+
+ odp = odp_lookup(ctx, offset * MLX5_IMR_MTT_SIZE,
+ nentries * MLX5_IMR_MTT_SIZE, mr);
+
+ for (i = 0; i < nentries; i++, pklm++) {
+ pklm->bcount = cpu_to_be32(MLX5_IMR_MTT_SIZE);
+ va = (offset + i) * MLX5_IMR_MTT_SIZE;
+ if (odp && odp->umem->address == va) {
+ struct mlx5_ib_mr *mtt = odp->private;
+
+ pklm->key = cpu_to_be32(mtt->ibmr.lkey);
+ odp = odp_next(odp);
+ } else {
+ pklm->key = cpu_to_be32(dev->null_mkey);
+ }
+ mlx5_ib_dbg(dev, "[%d] va %lx key %x\n",
+ i, va, be32_to_cpu(pklm->key));
+ }
+}
+
+static void mr_leaf_free_action(struct work_struct *work)
+{
+ struct ib_umem_odp *odp = container_of(work, struct ib_umem_odp, work);
+ int idx = ib_umem_start(odp->umem) >> MLX5_IMR_MTT_SHIFT;
+ struct mlx5_ib_mr *mr = odp->private, *imr = mr->parent;
+
+ mr->parent = NULL;
+ synchronize_srcu(&mr->dev->mr_srcu);
+
+ if (!READ_ONCE(odp->dying)) {
+ mr->parent = imr;
+ if (atomic_dec_and_test(&imr->num_leaf_free))
+ wake_up(&imr->q_leaf_free);
+ return;
+ }
+
+ ib_umem_release(odp->umem);
+ if (imr->live)
+ mlx5_ib_update_xlt(imr, idx, 1, 0,
+ MLX5_IB_UPD_XLT_INDIRECT |
+ MLX5_IB_UPD_XLT_ATOMIC);
+ mlx5_mr_cache_free(mr->dev, mr);
+
+ if (atomic_dec_and_test(&imr->num_leaf_free))
+ wake_up(&imr->q_leaf_free);
+}
+
void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start,
unsigned long end)
{
*/
ib_umem_odp_unmap_dma_pages(umem, start, end);
+
+ if (unlikely(!umem->npages && mr->parent &&
+ !umem->odp_data->dying)) {
+ WRITE_ONCE(umem->odp_data->dying, 1);
+ atomic_inc(&mr->parent->num_leaf_free);
+ schedule_work(&umem->odp_data->work);
+ }
}
void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic))
caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
+ if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) &&
+ MLX5_CAP_GEN(dev->mdev, null_mkey) &&
+ MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
+ caps->general_caps |= IB_ODP_SUPPORT_IMPLICIT;
+
return;
}
wq_num);
}
+static struct mlx5_ib_mr *implicit_mr_alloc(struct ib_pd *pd,
+ struct ib_umem *umem,
+ bool ksm, int access_flags)
+{
+ struct mlx5_ib_dev *dev = to_mdev(pd->device);
+ struct mlx5_ib_mr *mr;
+ int err;
+
+ mr = mlx5_mr_cache_alloc(dev, ksm ? MLX5_IMR_KSM_CACHE_ENTRY :
+ MLX5_IMR_MTT_CACHE_ENTRY);
+
+ if (IS_ERR(mr))
+ return mr;
+
+ mr->ibmr.pd = pd;
+
+ mr->dev = dev;
+ mr->access_flags = access_flags;
+ mr->mmkey.iova = 0;
+ mr->umem = umem;
+
+ if (ksm) {
+ err = mlx5_ib_update_xlt(mr, 0,
+ mlx5_imr_ksm_entries,
+ MLX5_KSM_PAGE_SHIFT,
+ MLX5_IB_UPD_XLT_INDIRECT |
+ MLX5_IB_UPD_XLT_ZAP |
+ MLX5_IB_UPD_XLT_ENABLE);
+
+ } else {
+ err = mlx5_ib_update_xlt(mr, 0,
+ MLX5_IMR_MTT_ENTRIES,
+ PAGE_SHIFT,
+ MLX5_IB_UPD_XLT_ZAP |
+ MLX5_IB_UPD_XLT_ENABLE |
+ MLX5_IB_UPD_XLT_ATOMIC);
+ }
+
+ if (err)
+ goto fail;
+
+ mr->ibmr.lkey = mr->mmkey.key;
+ mr->ibmr.rkey = mr->mmkey.key;
+
+ mr->live = 1;
+
+ mlx5_ib_dbg(dev, "key %x dev %p mr %p\n",
+ mr->mmkey.key, dev->mdev, mr);
+
+ return mr;
+
+fail:
+ mlx5_ib_err(dev, "Failed to register MKEY %d\n", err);
+ mlx5_mr_cache_free(dev, mr);
+
+ return ERR_PTR(err);
+}
+
+static struct ib_umem_odp *implicit_mr_get_data(struct mlx5_ib_mr *mr,
+ u64 io_virt, size_t bcnt)
+{
+ struct ib_ucontext *ctx = mr->ibmr.pd->uobject->context;
+ struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.pd->device);
+ struct ib_umem_odp *odp, *result = NULL;
+ u64 addr = io_virt & MLX5_IMR_MTT_MASK;
+ int nentries = 0, start_idx = 0, ret;
+ struct mlx5_ib_mr *mtt;
+ struct ib_umem *umem;
+
+ mutex_lock(&mr->umem->odp_data->umem_mutex);
+ odp = odp_lookup(ctx, addr, 1, mr);
+
+ mlx5_ib_dbg(dev, "io_virt:%llx bcnt:%zx addr:%llx odp:%p\n",
+ io_virt, bcnt, addr, odp);
+
+next_mr:
+ if (likely(odp)) {
+ if (nentries)
+ nentries++;
+ } else {
+ umem = ib_alloc_odp_umem(ctx, addr, MLX5_IMR_MTT_SIZE);
+ if (IS_ERR(umem)) {
+ mutex_unlock(&mr->umem->odp_data->umem_mutex);
+ return ERR_CAST(umem);
+ }
+
+ mtt = implicit_mr_alloc(mr->ibmr.pd, umem, 0, mr->access_flags);
+ if (IS_ERR(mtt)) {
+ mutex_unlock(&mr->umem->odp_data->umem_mutex);
+ ib_umem_release(umem);
+ return ERR_CAST(mtt);
+ }
+
+ odp = umem->odp_data;
+ odp->private = mtt;
+ mtt->umem = umem;
+ mtt->mmkey.iova = addr;
+ mtt->parent = mr;
+ INIT_WORK(&odp->work, mr_leaf_free_action);
+
+ if (!nentries)
+ start_idx = addr >> MLX5_IMR_MTT_SHIFT;
+ nentries++;
+ }
+
+ odp->dying = 0;
+
+ /* Return first odp if region not covered by single one */
+ if (likely(!result))
+ result = odp;
+
+ addr += MLX5_IMR_MTT_SIZE;
+ if (unlikely(addr < io_virt + bcnt)) {
+ odp = odp_next(odp);
+ if (odp && odp->umem->address != addr)
+ odp = NULL;
+ goto next_mr;
+ }
+
+ if (unlikely(nentries)) {
+ ret = mlx5_ib_update_xlt(mr, start_idx, nentries, 0,
+ MLX5_IB_UPD_XLT_INDIRECT |
+ MLX5_IB_UPD_XLT_ATOMIC);
+ if (ret) {
+ mlx5_ib_err(dev, "Failed to update PAS\n");
+ result = ERR_PTR(ret);
+ }
+ }
+
+ mutex_unlock(&mr->umem->odp_data->umem_mutex);
+ return result;
+}
+
+struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
+ int access_flags)
+{
+ struct ib_ucontext *ctx = pd->ibpd.uobject->context;
+ struct mlx5_ib_mr *imr;
+ struct ib_umem *umem;
+
+ umem = ib_umem_get(ctx, 0, 0, IB_ACCESS_ON_DEMAND, 0);
+ if (IS_ERR(umem))
+ return ERR_CAST(umem);
+
+ imr = implicit_mr_alloc(&pd->ibpd, umem, 1, access_flags);
+ if (IS_ERR(imr)) {
+ ib_umem_release(umem);
+ return ERR_CAST(imr);
+ }
+
+ imr->umem = umem;
+ init_waitqueue_head(&imr->q_leaf_free);
+ atomic_set(&imr->num_leaf_free, 0);
+
+ return imr;
+}
+
+static int mr_leaf_free(struct ib_umem *umem, u64 start,
+ u64 end, void *cookie)
+{
+ struct mlx5_ib_mr *mr = umem->odp_data->private, *imr = cookie;
+
+ if (mr->parent != imr)
+ return 0;
+
+ ib_umem_odp_unmap_dma_pages(umem,
+ ib_umem_start(umem),
+ ib_umem_end(umem));
+
+ if (umem->odp_data->dying)
+ return 0;
+
+ WRITE_ONCE(umem->odp_data->dying, 1);
+ atomic_inc(&imr->num_leaf_free);
+ schedule_work(&umem->odp_data->work);
+
+ return 0;
+}
+
+void mlx5_ib_free_implicit_mr(struct mlx5_ib_mr *imr)
+{
+ struct ib_ucontext *ctx = imr->ibmr.pd->uobject->context;
+
+ down_read(&ctx->umem_rwsem);
+ rbt_ib_umem_for_each_in_range(&ctx->umem_tree, 0, ULLONG_MAX,
+ mr_leaf_free, imr);
+ up_read(&ctx->umem_rwsem);
+
+ wait_event(imr->q_leaf_free, !atomic_read(&imr->num_leaf_free));
+}
+
/*
* Handle a single data segment in a page-fault WQE or RDMA region.
*
* -EFAULT when there's an error mapping the requested pages. The caller will
* abort the page fault handling.
*/
-static int pagefault_single_data_segment(struct mlx5_ib_dev *mib_dev,
+static int pagefault_single_data_segment(struct mlx5_ib_dev *dev,
u32 key, u64 io_virt, size_t bcnt,
u32 *bytes_committed,
u32 *bytes_mapped)
{
int srcu_key;
- unsigned int current_seq;
+ unsigned int current_seq = 0;
u64 start_idx;
int npages = 0, ret = 0;
struct mlx5_ib_mr *mr;
u64 access_mask = ODP_READ_ALLOWED_BIT;
+ struct ib_umem_odp *odp;
+ int implicit = 0;
+ size_t size;
- srcu_key = srcu_read_lock(&mib_dev->mr_srcu);
- mr = mlx5_ib_odp_find_mr_lkey(mib_dev, key);
+ srcu_key = srcu_read_lock(&dev->mr_srcu);
+ mr = mlx5_ib_odp_find_mr_lkey(dev, key);
/*
* If we didn't find the MR, it means the MR was closed while we were
* handling the ODP event. In this case we return -EFAULT so that the
* QP will be closed.
*/
if (!mr || !mr->ibmr.pd) {
- pr_err("Failed to find relevant mr for lkey=0x%06x, probably the MR was destroyed\n",
- key);
+ mlx5_ib_dbg(dev, "Failed to find relevant mr for lkey=0x%06x, probably the MR was destroyed\n",
+ key);
ret = -EFAULT;
goto srcu_unlock;
}
if (!mr->umem->odp_data) {
- pr_debug("skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
- key);
+ mlx5_ib_dbg(dev, "skipping non ODP MR (lkey=0x%06x) in page fault handler.\n",
+ key);
if (bytes_mapped)
*bytes_mapped +=
(bcnt - *bytes_committed);
goto srcu_unlock;
}
- current_seq = ACCESS_ONCE(mr->umem->odp_data->notifiers_seq);
- /*
- * Ensure the sequence number is valid for some time before we call
- * gup.
- */
- smp_rmb();
-
/*
* Avoid branches - this code will perform correctly
* in all iterations (in iteration 2 and above,
io_virt += *bytes_committed;
bcnt -= *bytes_committed;
+ if (!mr->umem->odp_data->page_list) {
+ odp = implicit_mr_get_data(mr, io_virt, bcnt);
+
+ if (IS_ERR(odp)) {
+ ret = PTR_ERR(odp);
+ goto srcu_unlock;
+ }
+ mr = odp->private;
+ implicit = 1;
+
+ } else {
+ odp = mr->umem->odp_data;
+ }
+
+next_mr:
+ current_seq = READ_ONCE(odp->notifiers_seq);
+ /*
+ * Ensure the sequence number is valid for some time before we call
+ * gup.
+ */
+ smp_rmb();
+
+ size = min_t(size_t, bcnt, ib_umem_end(odp->umem) - io_virt);
start_idx = (io_virt - (mr->mmkey.iova & PAGE_MASK)) >> PAGE_SHIFT;
if (mr->umem->writable)
access_mask |= ODP_WRITE_ALLOWED_BIT;
- npages = ib_umem_odp_map_dma_pages(mr->umem, io_virt, bcnt,
- access_mask, current_seq);
- if (npages < 0) {
- ret = npages;
+
+ ret = ib_umem_odp_map_dma_pages(mr->umem, io_virt, size,
+ access_mask, current_seq);
+
+ if (ret < 0)
goto srcu_unlock;
- }
- if (npages > 0) {
- mutex_lock(&mr->umem->odp_data->umem_mutex);
+ if (ret > 0) {
+ int np = ret;
+
+ mutex_lock(&odp->umem_mutex);
if (!ib_umem_mmu_notifier_retry(mr->umem, current_seq)) {
/*
* No need to check whether the MTTs really belong to
* this MR, since ib_umem_odp_map_dma_pages already
* checks this.
*/
- ret = mlx5_ib_update_xlt(mr, start_idx, npages,
+ ret = mlx5_ib_update_xlt(mr, start_idx, np,
PAGE_SHIFT,
MLX5_IB_UPD_XLT_ATOMIC);
} else {
ret = -EAGAIN;
}
- mutex_unlock(&mr->umem->odp_data->umem_mutex);
+ mutex_unlock(&odp->umem_mutex);
if (ret < 0) {
if (ret != -EAGAIN)
- pr_err("Failed to update mkey page tables\n");
+ mlx5_ib_err(dev, "Failed to update mkey page tables\n");
goto srcu_unlock;
}
if (bytes_mapped) {
- u32 new_mappings = npages * PAGE_SIZE -
+ u32 new_mappings = np * PAGE_SIZE -
(io_virt - round_down(io_virt, PAGE_SIZE));
- *bytes_mapped += min_t(u32, new_mappings, bcnt);
+ *bytes_mapped += min_t(u32, new_mappings, size);
}
+
+ npages += np;
+ }
+
+ bcnt -= size;
+ if (unlikely(bcnt)) {
+ struct ib_umem_odp *next;
+
+ io_virt += size;
+ next = odp_next(odp);
+ if (unlikely(!next || next->umem->address != io_virt)) {
+ mlx5_ib_dbg(dev, "next implicit leaf removed at 0x%llx. got %p\n",
+ io_virt, next);
+ ret = -EAGAIN;
+ goto srcu_unlock_no_wait;
+ }
+ odp = next;
+ mr = odp->private;
+ goto next_mr;
}
srcu_unlock:
if (ret == -EAGAIN) {
- if (!mr->umem->odp_data->dying) {
- struct ib_umem_odp *odp_data = mr->umem->odp_data;
+ if (implicit || !odp->dying) {
unsigned long timeout =
msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT);
if (!wait_for_completion_timeout(
- &odp_data->notifier_completion,
+ &odp->notifier_completion,
timeout)) {
- pr_warn("timeout waiting for mmu notifier completion\n");
+ mlx5_ib_warn(dev, "timeout waiting for mmu notifier. seq %d against %d\n",
+ current_seq, odp->notifiers_seq);
}
} else {
/* The MR is being killed, kill the QP as well. */
ret = -EFAULT;
}
}
- srcu_read_unlock(&mib_dev->mr_srcu, srcu_key);
+
+srcu_unlock_no_wait:
+ srcu_read_unlock(&dev->mr_srcu, srcu_key);
*bytes_committed = 0;
return ret ? ret : npages;
}
goto resolve_page_fault;
} else if (ret < 0 || total_wqe_bytes > bytes_mapped) {
if (ret != -ENOENT)
- mlx5_ib_err(dev, "Error getting user pages for page fault. Error: %d\n",
- ret);
+ mlx5_ib_err(dev, "PAGE FAULT error: %d. QP 0x%x. type: 0x%x\n",
+ ret, pfault->wqe.wq_num, pfault->type);
goto resolve_page_fault;
}
resolve_page_fault:
mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
- pfault->token, resume_with_error,
+ pfault->wqe.wq_num, resume_with_error,
pfault->type);
free_page((unsigned long)buffer);
}
ret = pagefault_single_data_segment(dev, rkey, address,
prefetch_len,
&bytes_committed, NULL);
- if (ret < 0) {
+ if (ret < 0 && ret != -EAGAIN) {
mlx5_ib_warn(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
- ret, pfault->token, address,
- prefetch_len);
+ ret, pfault->token, address, prefetch_len);
}
}
}
}
}
-int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev)
+void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
+{
+ if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
+ return;
+
+ switch (ent->order - 2) {
+ case MLX5_IMR_MTT_CACHE_ENTRY:
+ ent->page = PAGE_SHIFT;
+ ent->xlt = MLX5_IMR_MTT_ENTRIES *
+ sizeof(struct mlx5_mtt) /
+ MLX5_IB_UMR_OCTOWORD;
+ ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
+ ent->limit = 0;
+ break;
+
+ case MLX5_IMR_KSM_CACHE_ENTRY:
+ ent->page = MLX5_KSM_PAGE_SHIFT;
+ ent->xlt = mlx5_imr_ksm_entries *
+ sizeof(struct mlx5_klm) /
+ MLX5_IB_UMR_OCTOWORD;
+ ent->access_mode = MLX5_MKC_ACCESS_MODE_KSM;
+ ent->limit = 0;
+ break;
+ }
+}
+
+int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
{
int ret;
- ret = init_srcu_struct(&ibdev->mr_srcu);
+ ret = init_srcu_struct(&dev->mr_srcu);
if (ret)
return ret;
+ if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
+ ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
+ if (ret) {
+ mlx5_ib_err(dev, "Error getting null_mkey %d\n", ret);
+ return ret;
+ }
+ }
+
return 0;
}
-void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev)
+void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *dev)
+{
+ cleanup_srcu_struct(&dev->mr_srcu);
+}
+
+int mlx5_ib_odp_init(void)
{
- cleanup_srcu_struct(&ibdev->mr_srcu);
+ mlx5_imr_ksm_entries = BIT_ULL(get_order(TASK_SIZE) -
+ MLX5_IMR_MTT_BITS);
+
+ return 0;
}
projects / platform / kernel / linux-rpi.git / commitdiff