@@ -541,6 +541,7 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
struct mlx5_ib_ucontext *context;
struct mlx5_uuar_info *uuari;
struct mlx5_uar *uars;
+ int gross_uuars;
int num_uars;
int uuarn;
int err;
@@ -559,11 +560,13 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
if (req.total_num_uuars == 0)
return ERR_PTR(-EINVAL);
- req.total_num_uuars = ALIGN(req.total_num_uuars, MLX5_BF_REGS_PER_PAGE);
+ req.total_num_uuars = ALIGN(req.total_num_uuars,
+ MLX5_NON_FP_BF_REGS_PER_PAGE);
if (req.num_low_latency_uuars > req.total_num_uuars - 1)
return ERR_PTR(-EINVAL);
- num_uars = req.total_num_uuars / MLX5_BF_REGS_PER_PAGE;
+ num_uars = req.total_num_uuars / MLX5_NON_FP_BF_REGS_PER_PAGE;
+ gross_uuars = num_uars * MLX5_BF_REGS_PER_PAGE;
resp.qp_tab_size = 1 << dev->mdev.caps.log_max_qp;
resp.bf_reg_size = dev->mdev.caps.bf_reg_size;
resp.cache_line_size = L1_CACHE_BYTES;
@@ -585,7 +588,7 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
goto out_ctx;
}
- uuari->bitmap = kcalloc(BITS_TO_LONGS(req.total_num_uuars),
+ uuari->bitmap = kcalloc(BITS_TO_LONGS(gross_uuars),
sizeof(*uuari->bitmap),
GFP_KERNEL);
if (!uuari->bitmap) {
@@ -595,13 +598,13 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
/*
* clear all fast path uuars
*/
- for (i = 0; i < req.total_num_uuars; i++) {
+ for (i = 0; i < gross_uuars; i++) {
uuarn = i & 3;
if (uuarn == 2 || uuarn == 3)
set_bit(i, uuari->bitmap);
}
- uuari->count = kcalloc(req.total_num_uuars, sizeof(*uuari->count), GFP_KERNEL);
+ uuari->count = kcalloc(gross_uuars, sizeof(*uuari->count), GFP_KERNEL);
if (!uuari->count) {
err = -ENOMEM;
goto out_bitmap;
@@ -340,14 +340,57 @@ static int qp_has_rq(struct ib_qp_init_attr *attr)
return 1;
}
+static int first_med_uuar(void)
+{
+ return 1;
+}
+
+static int next_uuar(int n)
+{
+ n++;
+
+ while (((n % 4) & 2))
+ n++;
+
+ return n;
+}
+
+static int num_med_uuar(struct mlx5_uuar_info *uuari)
+{
+ int n;
+
+ n = uuari->num_uars * MLX5_NON_FP_BF_REGS_PER_PAGE -
+ uuari->num_low_latency_uuars - 1;
+
+ return n >= 0 ? n : 0;
+}
+
+static int max_uuari(struct mlx5_uuar_info *uuari)
+{
+ return uuari->num_uars * 4;
+}
+
+static int first_hi_uuar(struct mlx5_uuar_info *uuari)
+{
+ int med;
+ int i;
+ int t;
+
+ med = num_med_uuar(uuari);
+ for (t = 0, i = first_med_uuar();; i = next_uuar(i)) {
+ t++;
+ if (t == med)
+ return next_uuar(i);
+ }
+
+ return 0;
+}
+
static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
{
- int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
- int start_uuar;
int i;
- start_uuar = nuuars - uuari->num_low_latency_uuars;
- for (i = start_uuar; i < nuuars; i++) {
+ for (i = first_hi_uuar(uuari); i < max_uuari(uuari); i = next_uuar(i)) {
if (!test_bit(i, uuari->bitmap)) {
set_bit(i, uuari->bitmap);
uuari->count[i]++;
@@ -360,19 +403,10 @@ static int alloc_high_class_uuar(struct mlx5_uuar_info *uuari)
static int alloc_med_class_uuar(struct mlx5_uuar_info *uuari)
{
- int nuuars = uuari->num_uars * MLX5_BF_REGS_PER_PAGE;
- int minidx = 1;
- int uuarn;
- int end;
+ int minidx = first_med_uuar();
int i;
- end = nuuars - uuari->num_low_latency_uuars;
-
- for (i = 1; i < end; i++) {
- uuarn = i & 3;
- if (uuarn == 2 || uuarn == 3)
- continue;
-
+ for (i = first_med_uuar(); i < first_hi_uuar(uuari); i = next_uuar(i)) {
if (uuari->count[i] < uuari->count[minidx])
minidx = i;
}
@@ -510,11 +544,16 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_HIGH);
if (uuarn < 0) {
mlx5_ib_dbg(dev, "failed to allocate low latency UUAR\n");
- mlx5_ib_dbg(dev, "reverting to high latency\n");
- uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
+ mlx5_ib_dbg(dev, "reverting to medium latency\n");
+ uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_MEDIUM);
if (uuarn < 0) {
- mlx5_ib_dbg(dev, "uuar allocation failed\n");
- return uuarn;
+ mlx5_ib_dbg(dev, "failed to allocate medium latency UUAR\n");
+ mlx5_ib_dbg(dev, "reverting to high latency\n");
+ uuarn = alloc_uuar(&context->uuari, MLX5_IB_LATENCY_CLASS_LOW);
+ if (uuarn < 0) {
+ mlx5_ib_warn(dev, "uuar allocation failed\n");
+ return uuarn;
+ }
}
}
@@ -104,9 +104,10 @@ enum {
};
enum {
- MLX5_BF_REGS_PER_PAGE = 4,
- MLX5_MAX_UAR_PAGES = 1 << 8,
- MLX5_MAX_UUARS = MLX5_MAX_UAR_PAGES * MLX5_BF_REGS_PER_PAGE,
+ MLX5_BF_REGS_PER_PAGE = 4,
+ MLX5_MAX_UAR_PAGES = 1 << 8,
+ MLX5_NON_FP_BF_REGS_PER_PAGE = 2,
+ MLX5_MAX_UUARS = MLX5_MAX_UAR_PAGES * MLX5_NON_FP_BF_REGS_PER_PAGE,
};
enum {
The micro UAR (uuar) allocator had a bug which resulted from the fact that in each UAR we only have two micro UARs avaialable, those at index 0 and 1. This patch defines iterators to aid in traversing the list of available micro UARs when allocating a uuar. In addition, change the logic in create_user_qp() so that if high class allocation fails (high class means lower latency), we revert to medium class and not to the low class. Signed-off-by: Eli Cohen <eli@mellanox.com> --- drivers/infiniband/hw/mlx5/main.c | 13 ++++--- drivers/infiniband/hw/mlx5/qp.c | 77 +++++++++++++++++++++++++++++---------- include/linux/mlx5/device.h | 7 ++-- 3 files changed, 70 insertions(+), 27 deletions(-)