[PATCH 0/5] IB/ehca: SRQ and MR/MW fixes

[PATCH 0/5] IB/ehca: SRQ and MR/MW fixes

[Joachim Fenkes]

Here are some more fixes for the eHCA driver, fixing some problems we found
during internal system test.

[1/5] fixes the QP pointer determination for SRQ base QPs
[2/5] fixes a masking error in {,re}reg_phys_mr()
[3/5] fixes a bug in alloc_fmr() and simplifies some code
[4/5] refactors hca_cap_mr_pgsize and fixes a problem with ib_srp
[5/5] enables large page MRs by default

I built the patches on top of Roland's for-2.6.24 git branch. Please review
and queue them for 2.6.24-rc1 if you're okay with them. Thanks!

Cheers,
  Joachim

=2D-=20
Joachim Fenkes =A0-- =A0eHCA Linux Driver Developer and Hardware Tamer
IBM Deutschland Entwicklung GmbH =A0-- =A0Dept. 3627 (I/O Firmware Dev. 2)
Schoenaicher Strasse 220 =A0-- =A071032 Boeblingen =A0-- =A0Germany
eMail: fenkes@de.ibm.com

[PATCH 1/5] IB/ehca: Supply QP token for SRQ base QPs

[Joachim Fenkes]

Because hardware reports the SRQ token in RWQEs of SRQ base QPs, supply the
base QP token as SRQ token, so we can properly find the SRQ base QP.

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
---
 drivers/infiniband/hw/ehca/ehca_qp.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/drivers/infiniband/hw/ehca/ehca_qp.c b/drivers/infiniband/hw/ehca/ehca_qp.c
index e2bd62b..de18264 100644
--- a/drivers/infiniband/hw/ehca/ehca_qp.c
+++ b/drivers/infiniband/hw/ehca/ehca_qp.c
@@ -451,7 +451,6 @@ static struct ehca_qp *internal_create_qp(
 		has_srq = 1;
 		parms.ext_type = EQPT_SRQBASE;
 		parms.srq_qpn = my_srq->real_qp_num;
-		parms.srq_token = my_srq->token;
 	}
 
 	if (is_llqp && has_srq) {
@@ -583,6 +582,9 @@ static struct ehca_qp *internal_create_qp(
 		goto create_qp_exit1;
 	}
 
+	if (has_srq)
+		parms.srq_token = my_qp->token;
+
 	parms.servicetype = ibqptype2servicetype(qp_type);
 	if (parms.servicetype < 0) {
 		ret = -EINVAL;
-- 
1.5.2

[PATCH 2/5] IB/ehca: Fix masking error in {,re}reg_phys_mr()

[Joachim Fenkes]

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
---
 drivers/infiniband/hw/ehca/ehca_mrmw.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/hw/ehca/ehca_mrmw.c b/drivers/infiniband/hw/ehca/ehca_mrmw.c
index da88738..16c9efd 100644
--- a/drivers/infiniband/hw/ehca/ehca_mrmw.c
+++ b/drivers/infiniband/hw/ehca/ehca_mrmw.c
@@ -259,7 +259,7 @@ struct ib_mr *ehca_reg_phys_mr(struct ib_pd *pd,
 		pginfo.u.phy.num_phys_buf = num_phys_buf;
 		pginfo.u.phy.phys_buf_array = phys_buf_array;
 		pginfo.next_hwpage =
-			((u64)iova_start & ~(hw_pgsize - 1)) / hw_pgsize;
+			((u64)iova_start & ~PAGE_MASK) / hw_pgsize;
 
 		ret = ehca_reg_mr(shca, e_mr, iova_start, size, mr_access_flags,
 				  e_pd, &pginfo, &e_mr->ib.ib_mr.lkey,
@@ -547,7 +547,7 @@ int ehca_rereg_phys_mr(struct ib_mr *mr,
 		pginfo.u.phy.num_phys_buf = num_phys_buf;
 		pginfo.u.phy.phys_buf_array = phys_buf_array;
 		pginfo.next_hwpage =
-			((u64)iova_start & ~(hw_pgsize - 1)) / hw_pgsize;
+			((u64)iova_start & ~PAGE_MASK) / hw_pgsize;
 	}
 	if (mr_rereg_mask & IB_MR_REREG_ACCESS)
 		new_acl = mr_access_flags;
-- 
1.5.2

[PATCH 3/5] IB/ehca: Fix ehca_encode_hwpage_size() and alloc_fmr()

[Joachim Fenkes]

Simplify ehca_encode_hwpage_size(), fixing an infinite loop for pgsize == 0
in the process. Fix the bug in alloc_fmr() that triggered the loop.

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
---
 drivers/infiniband/hw/ehca/ehca_mrmw.c |   15 ++++-----------
 1 files changed, 4 insertions(+), 11 deletions(-)

diff --git a/drivers/infiniband/hw/ehca/ehca_mrmw.c b/drivers/infiniband/hw/ehca/ehca_mrmw.c
index 16c9efd..b9a788c 100644
--- a/drivers/infiniband/hw/ehca/ehca_mrmw.c
+++ b/drivers/infiniband/hw/ehca/ehca_mrmw.c
@@ -72,17 +72,9 @@ enum ehca_mr_pgsize {
 
 static u32 ehca_encode_hwpage_size(u32 pgsize)
 {
-	u32 idx = 0;
-	pgsize >>= 12;
-	/*
-	 * map mr page size into hw code:
-	 * 0, 1, 2, 3 for 4K, 64K, 1M, 64M
-	 */
-	while (!(pgsize & 1)) {
-		idx++;
-		pgsize >>= 4;
-	}
-	return idx;
+	int log = ilog2(pgsize);
+	WARN_ON(log < 12 || log > 24 || log & 3);
+	return (log - 12) / 4;
 }
 
 static u64 ehca_get_max_hwpage_size(struct ehca_shca *shca)
@@ -826,6 +818,7 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
 
 	/* register MR on HCA */
 	memset(&pginfo, 0, sizeof(pginfo));
+	pginfo.hwpage_size = hw_pgsize;
 	/*
 	 * pginfo.num_hwpages==0, ie register_rpages() will not be called
 	 * but deferred to map_phys_fmr()
-- 
1.5.2

[PATCH 4/5] IB/ehca: Change meaning of hca_cap_mr_pgsize

[Joachim Fenkes]

ehca_shca.hca_cap_mr_pgsize now contains all supported page sizes ORed
together. This makes some checks easier to code and understand, plus we can
return this value verbatim in query_hca(), fixing a problem with SRP
(reported by Anton Blanchard -- thanks!).

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
---
 drivers/infiniband/hw/ehca/ehca_classes.h |    1 -
 drivers/infiniband/hw/ehca/ehca_hca.c     |    1 +
 drivers/infiniband/hw/ehca/ehca_main.c    |   18 ++++++++++++-
 drivers/infiniband/hw/ehca/ehca_mrmw.c    |   38 ++++++++++++++--------------
 4 files changed, 36 insertions(+), 22 deletions(-)

diff --git a/drivers/infiniband/hw/ehca/ehca_classes.h b/drivers/infiniband/hw/ehca/ehca_classes.h
index 0f7a55d..365bc5d 100644
--- a/drivers/infiniband/hw/ehca/ehca_classes.h
+++ b/drivers/infiniband/hw/ehca/ehca_classes.h
@@ -323,7 +323,6 @@ extern int ehca_static_rate;
 extern int ehca_port_act_time;
 extern int ehca_use_hp_mr;
 extern int ehca_scaling_code;
-extern int ehca_mr_largepage;
 
 struct ipzu_queue_resp {
 	u32 qe_size;      /* queue entry size */
diff --git a/drivers/infiniband/hw/ehca/ehca_hca.c b/drivers/infiniband/hw/ehca/ehca_hca.c
index 4aa3ffa..15806d1 100644
--- a/drivers/infiniband/hw/ehca/ehca_hca.c
+++ b/drivers/infiniband/hw/ehca/ehca_hca.c
@@ -77,6 +77,7 @@ int ehca_query_device(struct ib_device *ibdev, struct ib_device_attr *props)
 	}
 
 	memset(props, 0, sizeof(struct ib_device_attr));
+	props->page_size_cap   = shca->hca_cap_mr_pgsize;
 	props->fw_ver          = rblock->hw_ver;
 	props->max_mr_size     = rblock->max_mr_size;
 	props->vendor_id       = rblock->vendor_id >> 8;
diff --git a/drivers/infiniband/hw/ehca/ehca_main.c b/drivers/infiniband/hw/ehca/ehca_main.c
index 403467f..d477dc3 100644
--- a/drivers/infiniband/hw/ehca/ehca_main.c
+++ b/drivers/infiniband/hw/ehca/ehca_main.c
@@ -260,13 +260,20 @@ static struct cap_descr {
 	{ HCA_CAP_MINI_QP, "HCA_CAP_MINI_QP" },
 };
 
-int ehca_sense_attributes(struct ehca_shca *shca)
+static int ehca_sense_attributes(struct ehca_shca *shca)
 {
 	int i, ret = 0;
 	u64 h_ret;
 	struct hipz_query_hca *rblock;
 	struct hipz_query_port *port;
 
+	static const u32 pgsize_map[] = {
+		HCA_CAP_MR_PGSIZE_4K,  0x1000,
+		HCA_CAP_MR_PGSIZE_64K, 0x10000,
+		HCA_CAP_MR_PGSIZE_1M,  0x100000,
+		HCA_CAP_MR_PGSIZE_16M, 0x1000000,
+	};
+
 	rblock = ehca_alloc_fw_ctrlblock(GFP_KERNEL);
 	if (!rblock) {
 		ehca_gen_err("Cannot allocate rblock memory.");
@@ -329,8 +336,15 @@ int ehca_sense_attributes(struct ehca_shca *shca)
 		if (EHCA_BMASK_GET(hca_cap_descr[i].mask, shca->hca_cap))
 			ehca_gen_dbg("   %s", hca_cap_descr[i].descr);
 
-	shca->hca_cap_mr_pgsize = rblock->memory_page_size_supported;
+	/* translate supported MR page sizes; always support 4K */
+	shca->hca_cap_mr_pgsize = EHCA_PAGESIZE;
+	if (ehca_mr_largepage) { /* support extra sizes only if enabled */
+		for (i = 0; i < ARRAY_SIZE(pgsize_map); i += 2)
+			if (rblock->memory_page_size_supported & pgsize_map[i])
+				shca->hca_cap_mr_pgsize |= pgsize_map[i + 1];
+	}
 
+	/* query max MTU from first port -- it's the same for all ports */
 	port = (struct hipz_query_port *)rblock;
 	h_ret = hipz_h_query_port(shca->ipz_hca_handle, 1, port);
 	if (h_ret != H_SUCCESS) {
diff --git a/drivers/infiniband/hw/ehca/ehca_mrmw.c b/drivers/infiniband/hw/ehca/ehca_mrmw.c
index b9a788c..bb97915 100644
--- a/drivers/infiniband/hw/ehca/ehca_mrmw.c
+++ b/drivers/infiniband/hw/ehca/ehca_mrmw.c
@@ -79,9 +79,7 @@ static u32 ehca_encode_hwpage_size(u32 pgsize)
 
 static u64 ehca_get_max_hwpage_size(struct ehca_shca *shca)
 {
-	if (shca->hca_cap_mr_pgsize & HCA_CAP_MR_PGSIZE_16M)
-		return EHCA_MR_PGSIZE16M;
-	return EHCA_MR_PGSIZE4K;
+	return 1UL << ilog2(shca->hca_cap_mr_pgsize);
 }
 
 static struct ehca_mr *ehca_mr_new(void)
@@ -288,7 +286,7 @@ struct ib_mr *ehca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 		container_of(pd->device, struct ehca_shca, ib_device);
 	struct ehca_pd *e_pd = container_of(pd, struct ehca_pd, ib_pd);
 	struct ehca_mr_pginfo pginfo;
-	int ret;
+	int ret, page_shift;
 	u32 num_kpages;
 	u32 num_hwpages;
 	u64 hwpage_size;
@@ -343,19 +341,20 @@ struct ib_mr *ehca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	/* determine number of MR pages */
 	num_kpages = NUM_CHUNKS((virt % PAGE_SIZE) + length, PAGE_SIZE);
 	/* select proper hw_pgsize */
-	if (ehca_mr_largepage &&
-	    (shca->hca_cap_mr_pgsize & HCA_CAP_MR_PGSIZE_16M)) {
-		int page_shift = PAGE_SHIFT;
-		if (e_mr->umem->hugetlb) {
-			/* determine page_shift, clamp between 4K and 16M */
-			page_shift = (fls64(length - 1) + 3) & ~3;
-			page_shift = min(max(page_shift, EHCA_MR_PGSHIFT4K),
-					 EHCA_MR_PGSHIFT16M);
-		}
-		hwpage_size = 1UL << page_shift;
-	} else
-		hwpage_size = EHCA_MR_PGSIZE4K; /* ehca1 only supports 4k */
-	ehca_dbg(pd->device, "hwpage_size=%lx", hwpage_size);
+	page_shift = PAGE_SHIFT;
+	if (e_mr->umem->hugetlb) {
+		/* determine page_shift, clamp between 4K and 16M */
+		page_shift = (fls64(length - 1) + 3) & ~3;
+		page_shift = min(max(page_shift, EHCA_MR_PGSHIFT4K),
+				 EHCA_MR_PGSHIFT16M);
+	}
+	hwpage_size = 1UL << page_shift;
+
+	/* now that we have the desired page size, shift until it's
+	 * supported, too. 4K is always supported, so this terminates.
+	 */
+	while (!(hwpage_size & shca->hca_cap_mr_pgsize))
+		hwpage_size >>= 4;
 
 reg_user_mr_fallback:
 	num_hwpages = NUM_CHUNKS((virt % hwpage_size) + length, hwpage_size);
@@ -801,8 +800,9 @@ struct ib_fmr *ehca_alloc_fmr(struct ib_pd *pd,
 		ib_fmr = ERR_PTR(-EINVAL);
 		goto alloc_fmr_exit0;
 	}
-	hw_pgsize = ehca_get_max_hwpage_size(shca);
-	if ((1 << fmr_attr->page_shift) != hw_pgsize) {
+
+	hw_pgsize = 1 << fmr_attr->page_shift;
+	if (!(hw_pgsize & shca->hca_cap_mr_pgsize)) {
 		ehca_err(pd->device, "unsupported fmr_attr->page_shift=%x",
 			 fmr_attr->page_shift);
 		ib_fmr = ERR_PTR(-EINVAL);
-- 
1.5.2

[PATCH 5/5] IB/ehca: Enable large page MRs by default

[Joachim Fenkes]

Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
---
 drivers/infiniband/hw/ehca/ehca_main.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/drivers/infiniband/hw/ehca/ehca_main.c b/drivers/infiniband/hw/ehca/ehca_main.c
index d477dc3..2f51c13 100644
--- a/drivers/infiniband/hw/ehca/ehca_main.c
+++ b/drivers/infiniband/hw/ehca/ehca_main.c
@@ -65,7 +65,7 @@ int ehca_port_act_time = 30;
 int ehca_poll_all_eqs  = 1;
 int ehca_static_rate   = -1;
 int ehca_scaling_code  = 0;
-int ehca_mr_largepage  = 0;
+int ehca_mr_largepage  = 1;
 
 module_param_named(open_aqp1,     ehca_open_aqp1,     int, S_IRUGO);
 module_param_named(debug_level,   ehca_debug_level,   int, S_IRUGO);
-- 
1.5.2

Re: [PATCH 5/5] IB/ehca: Enable large page MRs by default

[Roland Dreier]

thanks, applied 1-5