[PATCH for-next v3 0/4] Add support for 0xefa1 device

[PATCH for-next v3 0/4] Add support for 0xefa1 device

[Gal Pressman]

Hi all,

The following submission adds the needed functionality in order to
support 0xefa1 devices, and adds it to the driver pci table.

PR was sent:
https://github.com/linux-rdma/rdma-core/pull/789

Changelog -
v2->v3: https://lore.kernel.org/linux-rdma/20200720080113.13055-1-galpress@amazon.com/
* Remove gcc specific stuff to fix clang compilation.
v1->v2: https://lore.kernel.org/linux-rdma/20200709083630.21377-1-galpress@amazon.com/
* Add handshake with userspace provider to verify required features
  support.

Regards,
Gal

Gal Pressman (4):
  RDMA/efa: Expose maximum TX doorbell batch
  RDMA/efa: Expose minimum SQ size
  RDMA/efa: User/kernel compatibility handshake mechanism
  RDMA/efa: Add EFA 0xefa1 PCI ID

 .../infiniband/hw/efa/efa_admin_cmds_defs.h   | 15 +++++-
 drivers/infiniband/hw/efa/efa_com_cmd.c       |  2 +
 drivers/infiniband/hw/efa/efa_com_cmd.h       |  2 +
 drivers/infiniband/hw/efa/efa_main.c          |  6 ++-
 drivers/infiniband/hw/efa/efa_verbs.c         | 51 +++++++++++++++++++
 include/uapi/rdma/efa-abi.h                   | 15 +++++-
 6 files changed, 86 insertions(+), 5 deletions(-)


base-commit: 5f0b2a6093a4d9aab093964c65083fe801ef1e58
-- 
2.27.0

[PATCH for-next v3 1/4] RDMA/efa: Expose maximum TX doorbell batch

[Gal Pressman]

The device reports the maximum number of bytes to be written before
ringing the doorbell (zero means unlimited).

This patch queries the max batch size and reports it back to the
userspace library.

Reviewed-by: Daniel Kranzdorf <dkkranzd@amazon.com>
Reviewed-by: Firas JahJah <firasj@amazon.com>
Signed-off-by: Gal Pressman <galpress@amazon.com>
---
 drivers/infiniband/hw/efa/efa_admin_cmds_defs.h | 11 +++++++++++
 drivers/infiniband/hw/efa/efa_com_cmd.c         |  1 +
 drivers/infiniband/hw/efa/efa_com_cmd.h         |  1 +
 drivers/infiniband/hw/efa/efa_verbs.c           |  1 +
 include/uapi/rdma/efa-abi.h                     |  4 +++-
 5 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
index bef2bd291054..03e7388af06e 100644
--- a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
+++ b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
@@ -632,6 +632,17 @@ struct efa_admin_feature_queue_attr_desc {
 
 	/* Maximum number of SGEs for a single RDMA read WQE */
 	u16 max_wr_rdma_sges;
+
+	/*
+	 * Maximum number of bytes that can be written to SQ between two
+	 * consecutive doorbells (in units of 64B). Driver must ensure that only
+	 * complete WQEs are written to queue before issuing a doorbell.
+	 * Examples: max_tx_batch=16 and WQE size = 64B, means up to 16 WQEs can
+	 * be written to SQ between two consecutive doorbells. max_tx_batch=11
+	 * and WQE size = 128B, means up to 5 WQEs can be written to SQ between
+	 * two consecutive doorbells. Zero means unlimited.
+	 */
+	u16 max_tx_batch;
 };
 
 struct efa_admin_feature_aenq_desc {
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.c b/drivers/infiniband/hw/efa/efa_com_cmd.c
index fabd8df2e78f..53cfde5c43d8 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.c
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.c
@@ -480,6 +480,7 @@ int efa_com_get_device_attr(struct efa_com_dev *edev,
 	result->max_llq_size = resp.u.queue_attr.max_llq_size;
 	result->sub_cqs_per_cq = resp.u.queue_attr.sub_cqs_per_cq;
 	result->max_wr_rdma_sge = resp.u.queue_attr.max_wr_rdma_sges;
+	result->max_tx_batch = resp.u.queue_attr.max_tx_batch;
 
 	err = efa_com_get_feature(edev, &resp, EFA_ADMIN_NETWORK_ATTR);
 	if (err) {
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.h b/drivers/infiniband/hw/efa/efa_com_cmd.h
index 41ce4a476ee6..8df2a26d57d4 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.h
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.h
@@ -127,6 +127,7 @@ struct efa_com_get_device_attr_result {
 	u16 max_sq_sge;
 	u16 max_rq_sge;
 	u16 max_wr_rdma_sge;
+	u16 max_tx_batch;
 	u8 db_bar;
 };
 
diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
index 08313f7c73bc..f49d14cebe4a 100644
--- a/drivers/infiniband/hw/efa/efa_verbs.c
+++ b/drivers/infiniband/hw/efa/efa_verbs.c
@@ -1525,6 +1525,7 @@ int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
 	resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
 	resp.inline_buf_size = dev->dev_attr.inline_buf_size;
 	resp.max_llq_size = dev->dev_attr.max_llq_size;
+	resp.max_tx_batch = dev->dev_attr.max_tx_batch;
 
 	if (udata && udata->outlen) {
 		err = ib_copy_to_udata(udata, &resp,
diff --git a/include/uapi/rdma/efa-abi.h b/include/uapi/rdma/efa-abi.h
index 53b6e2036a9b..10781763da37 100644
--- a/include/uapi/rdma/efa-abi.h
+++ b/include/uapi/rdma/efa-abi.h
@@ -1,6 +1,6 @@
 /* SPDX-License-Identifier: ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) */
 /*
- * Copyright 2018-2019 Amazon.com, Inc. or its affiliates. All rights reserved.
+ * Copyright 2018-2020 Amazon.com, Inc. or its affiliates. All rights reserved.
  */
 
 #ifndef EFA_ABI_USER_H
@@ -31,6 +31,8 @@ struct efa_ibv_alloc_ucontext_resp {
 	__u16 sub_cqs_per_cq;
 	__u16 inline_buf_size;
 	__u32 max_llq_size; /* bytes */
+	__u16 max_tx_batch; /* units of 64 bytes */
+	__u8 reserved_90[6];
 };
 
 struct efa_ibv_alloc_pd_resp {
-- 
2.27.0

[PATCH for-next v3 2/4] RDMA/efa: Expose minimum SQ size

[Gal Pressman]

The device reports the minimum SQ size required for creation.

This patch queries the min SQ size and reports it back to the
userspace library.

Reviewed-by: Firas JahJah <firasj@amazon.com>
Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
Signed-off-by: Gal Pressman <galpress@amazon.com>
---
 drivers/infiniband/hw/efa/efa_admin_cmds_defs.h | 4 ++--
 drivers/infiniband/hw/efa/efa_com_cmd.c         | 1 +
 drivers/infiniband/hw/efa/efa_com_cmd.h         | 1 +
 drivers/infiniband/hw/efa/efa_verbs.c           | 1 +
 include/uapi/rdma/efa-abi.h                     | 3 ++-
 5 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
index 03e7388af06e..5484b08bbc5d 100644
--- a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
+++ b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
@@ -606,8 +606,8 @@ struct efa_admin_feature_queue_attr_desc {
 	/* Number of sub-CQs to be created for each CQ */
 	u16 sub_cqs_per_cq;
 
-	/* MBZ */
-	u16 reserved;
+	/* Minimum number of WQEs per SQ */
+	u16 min_sq_depth;
 
 	/* Maximum number of SGEs (buffers) allowed for a single send WQE */
 	u16 max_wr_send_sges;
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.c b/drivers/infiniband/hw/efa/efa_com_cmd.c
index 53cfde5c43d8..6ac23627f65a 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.c
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.c
@@ -481,6 +481,7 @@ int efa_com_get_device_attr(struct efa_com_dev *edev,
 	result->sub_cqs_per_cq = resp.u.queue_attr.sub_cqs_per_cq;
 	result->max_wr_rdma_sge = resp.u.queue_attr.max_wr_rdma_sges;
 	result->max_tx_batch = resp.u.queue_attr.max_tx_batch;
+	result->min_sq_depth = resp.u.queue_attr.min_sq_depth;
 
 	err = efa_com_get_feature(edev, &resp, EFA_ADMIN_NETWORK_ATTR);
 	if (err) {
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.h b/drivers/infiniband/hw/efa/efa_com_cmd.h
index 8df2a26d57d4..190bac23f585 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.h
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.h
@@ -128,6 +128,7 @@ struct efa_com_get_device_attr_result {
 	u16 max_rq_sge;
 	u16 max_wr_rdma_sge;
 	u16 max_tx_batch;
+	u16 min_sq_depth;
 	u8 db_bar;
 };
 
diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
index f49d14cebe4a..26102ab333b2 100644
--- a/drivers/infiniband/hw/efa/efa_verbs.c
+++ b/drivers/infiniband/hw/efa/efa_verbs.c
@@ -1526,6 +1526,7 @@ int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
 	resp.inline_buf_size = dev->dev_attr.inline_buf_size;
 	resp.max_llq_size = dev->dev_attr.max_llq_size;
 	resp.max_tx_batch = dev->dev_attr.max_tx_batch;
+	resp.min_sq_wr = dev->dev_attr.min_sq_depth;
 
 	if (udata && udata->outlen) {
 		err = ib_copy_to_udata(udata, &resp,
diff --git a/include/uapi/rdma/efa-abi.h b/include/uapi/rdma/efa-abi.h
index 10781763da37..7ef2306f8dd4 100644
--- a/include/uapi/rdma/efa-abi.h
+++ b/include/uapi/rdma/efa-abi.h
@@ -32,7 +32,8 @@ struct efa_ibv_alloc_ucontext_resp {
 	__u16 inline_buf_size;
 	__u32 max_llq_size; /* bytes */
 	__u16 max_tx_batch; /* units of 64 bytes */
-	__u8 reserved_90[6];
+	__u16 min_sq_wr;
+	__u8 reserved_a0[4];
 };
 
 struct efa_ibv_alloc_pd_resp {
-- 
2.27.0

[PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Gal Pressman]

Introduce a mechanism that performs an handshake between the userspace
provider and kernel driver which verifies that the user supports all
required features in order to operate correctly.

The handshake verifies the needed functionality by comparing the
reported device caps and the provider caps. If the device reports a
non-zero capability the appropriate comp mask is required from the
userspace provider in order to allocate the context.

Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
Signed-off-by: Gal Pressman <galpress@amazon.com>
---
 drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
 include/uapi/rdma/efa-abi.h           | 10 ++++++
 2 files changed, 59 insertions(+)

diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
index 26102ab333b2..7ca40df81ee5 100644
--- a/drivers/infiniband/hw/efa/efa_verbs.c
+++ b/drivers/infiniband/hw/efa/efa_verbs.c
@@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
 	return efa_com_dealloc_uar(&dev->edev, &params);
 }
 
+#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
+	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
+
+#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
+	{                                                                      \
+		.attr = #_attr,                                                \
+		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
+	}
+
+int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
+			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
+{
+	struct efa_dev *dev = to_edev(ibucontext->device);
+	int i;
+	struct {
+		char *attr;
+		bool check_comp;
+	} user_comp_handshakes[] = {
+		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
+				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
+		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
+				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
+	};
+
+	for (i = 0; i < ARRAY_SIZE(user_comp_handshakes); i++) {
+		if (!user_comp_handshakes[i].check_comp) {
+			ibdev_dbg(&dev->ibdev,
+				  "Userspace handshake failed for %s attribute\n",
+				  user_comp_handshakes[i].attr);
+			return -EOPNOTSUPP;
+		}
+	}
+
+	return 0;
+}
+
 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
 {
 	struct efa_ucontext *ucontext = to_eucontext(ibucontext);
 	struct efa_dev *dev = to_edev(ibucontext->device);
 	struct efa_ibv_alloc_ucontext_resp resp = {};
+	struct efa_ibv_alloc_ucontext_cmd cmd = {};
 	struct efa_com_alloc_uar_result result;
 	int err;
 
@@ -1514,6 +1551,18 @@ int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
 	 * we will ack input fields in our response.
 	 */
 
+	err = ib_copy_from_udata(&cmd, udata,
+				 min(sizeof(cmd), udata->inlen));
+	if (err) {
+		ibdev_dbg(&dev->ibdev,
+			  "Cannot copy udata for alloc_ucontext\n");
+		goto err_out;
+	}
+
+	err = efa_user_comp_handshake(ibucontext, &cmd);
+	if (err)
+		goto err_out;
+
 	err = efa_com_alloc_uar(&dev->edev, &result);
 	if (err)
 		goto err_out;
diff --git a/include/uapi/rdma/efa-abi.h b/include/uapi/rdma/efa-abi.h
index 7ef2306f8dd4..507a2862bedb 100644
--- a/include/uapi/rdma/efa-abi.h
+++ b/include/uapi/rdma/efa-abi.h
@@ -20,6 +20,16 @@
  * hex bit offset of the field.
  */
 
+enum {
+	EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH  = 1 << 0,
+	EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR = 1 << 1,
+};
+
+struct efa_ibv_alloc_ucontext_cmd {
+	__u32 comp_mask;
+	__u8 reserved_20[4];
+};
+
 enum efa_ibv_user_cmds_supp_udata {
 	EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE = 1 << 0,
 	EFA_USER_CMDS_SUPP_UDATA_CREATE_AH    = 1 << 1,
-- 
2.27.0

Re: [PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Jason Gunthorpe]

On Tue, Jul 21, 2020 at 04:30:48PM +0300, Gal Pressman wrote:
> Introduce a mechanism that performs an handshake between the userspace
> provider and kernel driver which verifies that the user supports all
> required features in order to operate correctly.
> 
> The handshake verifies the needed functionality by comparing the
> reported device caps and the provider caps. If the device reports a
> non-zero capability the appropriate comp mask is required from the
> userspace provider in order to allocate the context.
> 
> Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
> Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
> Signed-off-by: Gal Pressman <galpress@amazon.com>
>  drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
>  include/uapi/rdma/efa-abi.h           | 10 ++++++
>  2 files changed, 59 insertions(+)
> 
> diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
> index 26102ab333b2..7ca40df81ee5 100644
> +++ b/drivers/infiniband/hw/efa/efa_verbs.c
> @@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
>  	return efa_com_dealloc_uar(&dev->edev, &params);
>  }
>  
> +#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
> +	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
> +
> +#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
> +	{                                                                      \
> +		.attr = #_attr,                                                \
> +		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
> +	}
> +
> +int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
> +			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
> +{
> +	struct efa_dev *dev = to_edev(ibucontext->device);
> +	int i;
> +	struct {
> +		char *attr;
> +		bool check_comp;
> +	} user_comp_handshakes[] = {
> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
> +	};

This seems like a very expensive construct

Why have the array at all? Just list the macros and have them jump to
err

Jason

Re: [PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Gal Pressman]

On 22/07/2020 14:55, Jason Gunthorpe wrote:
> On Tue, Jul 21, 2020 at 04:30:48PM +0300, Gal Pressman wrote:
>> Introduce a mechanism that performs an handshake between the userspace
>> provider and kernel driver which verifies that the user supports all
>> required features in order to operate correctly.
>>
>> The handshake verifies the needed functionality by comparing the
>> reported device caps and the provider caps. If the device reports a
>> non-zero capability the appropriate comp mask is required from the
>> userspace provider in order to allocate the context.
>>
>> Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
>> Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
>> Signed-off-by: Gal Pressman <galpress@amazon.com>
>>  drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
>>  include/uapi/rdma/efa-abi.h           | 10 ++++++
>>  2 files changed, 59 insertions(+)
>>
>> diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
>> index 26102ab333b2..7ca40df81ee5 100644
>> +++ b/drivers/infiniband/hw/efa/efa_verbs.c
>> @@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
>>  	return efa_com_dealloc_uar(&dev->edev, &params);
>>  }
>>  
>> +#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
>> +	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
>> +
>> +#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
>> +	{                                                                      \
>> +		.attr = #_attr,                                                \
>> +		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
>> +	}
>> +
>> +int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
>> +			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
>> +{
>> +	struct efa_dev *dev = to_edev(ibucontext->device);
>> +	int i;
>> +	struct {
>> +		char *attr;
>> +		bool check_comp;
>> +	} user_comp_handshakes[] = {
>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
>> +	};
> 
> This seems like a very expensive construct
> 
> Why have the array at all? Just list the macros and have them jump to
> err

Do you mean:

if (CHECK_COMP(x1)) {
    ibdev_dbg(err);
    goto err;
}

if (CHECK_COMP(x2)) {
    ibdev_dbg(err);
    goto err;
}

[...]

That adds much more boilerplate code for each feature. Or do you have something
else in mind?

Re: [PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Jason Gunthorpe]

On Wed, Jul 22, 2020 at 03:04:20PM +0300, Gal Pressman wrote:
> On 22/07/2020 14:55, Jason Gunthorpe wrote:
> > On Tue, Jul 21, 2020 at 04:30:48PM +0300, Gal Pressman wrote:
> >> Introduce a mechanism that performs an handshake between the userspace
> >> provider and kernel driver which verifies that the user supports all
> >> required features in order to operate correctly.
> >>
> >> The handshake verifies the needed functionality by comparing the
> >> reported device caps and the provider caps. If the device reports a
> >> non-zero capability the appropriate comp mask is required from the
> >> userspace provider in order to allocate the context.
> >>
> >> Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
> >> Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
> >> Signed-off-by: Gal Pressman <galpress@amazon.com>
> >>  drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
> >>  include/uapi/rdma/efa-abi.h           | 10 ++++++
> >>  2 files changed, 59 insertions(+)
> >>
> >> diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
> >> index 26102ab333b2..7ca40df81ee5 100644
> >> +++ b/drivers/infiniband/hw/efa/efa_verbs.c
> >> @@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
> >>  	return efa_com_dealloc_uar(&dev->edev, &params);
> >>  }
> >>  
> >> +#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
> >> +	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
> >> +
> >> +#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
> >> +	{                                                                      \
> >> +		.attr = #_attr,                                                \
> >> +		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
> >> +	}
> >> +
> >> +int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
> >> +			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
> >> +{
> >> +	struct efa_dev *dev = to_edev(ibucontext->device);
> >> +	int i;
> >> +	struct {
> >> +		char *attr;
> >> +		bool check_comp;
> >> +	} user_comp_handshakes[] = {
> >> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
> >> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
> >> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
> >> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
> >> +	};
> > 
> > This seems like a very expensive construct
> > 
> > Why have the array at all? Just list the macros and have them jump to
> > err
> 
> Do you mean:
> 
> if (CHECK_COMP(x1)) {
>     ibdev_dbg(err);
>     goto err;
> }
> 
> if (CHECK_COMP(x2)) {
>     ibdev_dbg(err);
>     goto err;
> }
> 
> [...]
> 
> That adds much more boilerplate code for each feature. Or do you have something
> else in mind?

#define DO_COMP_HANDSHAKE() \
    if (...) goto err

DO_COMP_HANDSHAKE(x1)
DO_COMP_HANDSHAKE(2)

Jason

Re: [PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Gal Pressman]

On 22/07/2020 15:08, Jason Gunthorpe wrote:
> On Wed, Jul 22, 2020 at 03:04:20PM +0300, Gal Pressman wrote:
>> On 22/07/2020 14:55, Jason Gunthorpe wrote:
>>> On Tue, Jul 21, 2020 at 04:30:48PM +0300, Gal Pressman wrote:
>>>> Introduce a mechanism that performs an handshake between the userspace
>>>> provider and kernel driver which verifies that the user supports all
>>>> required features in order to operate correctly.
>>>>
>>>> The handshake verifies the needed functionality by comparing the
>>>> reported device caps and the provider caps. If the device reports a
>>>> non-zero capability the appropriate comp mask is required from the
>>>> userspace provider in order to allocate the context.
>>>>
>>>> Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
>>>> Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
>>>> Signed-off-by: Gal Pressman <galpress@amazon.com>
>>>>  drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
>>>>  include/uapi/rdma/efa-abi.h           | 10 ++++++
>>>>  2 files changed, 59 insertions(+)
>>>>
>>>> diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
>>>> index 26102ab333b2..7ca40df81ee5 100644
>>>> +++ b/drivers/infiniband/hw/efa/efa_verbs.c
>>>> @@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
>>>>  	return efa_com_dealloc_uar(&dev->edev, &params);
>>>>  }
>>>>  
>>>> +#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
>>>> +	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
>>>> +
>>>> +#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
>>>> +	{                                                                      \
>>>> +		.attr = #_attr,                                                \
>>>> +		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
>>>> +	}
>>>> +
>>>> +int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
>>>> +			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
>>>> +{
>>>> +	struct efa_dev *dev = to_edev(ibucontext->device);
>>>> +	int i;
>>>> +	struct {
>>>> +		char *attr;
>>>> +		bool check_comp;
>>>> +	} user_comp_handshakes[] = {
>>>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
>>>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
>>>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
>>>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
>>>> +	};
>>>
>>> This seems like a very expensive construct
>>>
>>> Why have the array at all? Just list the macros and have them jump to
>>> err
>>
>> Do you mean:
>>
>> if (CHECK_COMP(x1)) {
>>     ibdev_dbg(err);
>>     goto err;
>> }
>>
>> if (CHECK_COMP(x2)) {
>>     ibdev_dbg(err);
>>     goto err;
>> }
>>
>> [...]
>>
>> That adds much more boilerplate code for each feature. Or do you have something
>> else in mind?
> 
> #define DO_COMP_HANDSHAKE() \
>     if (...) goto err
> 
> DO_COMP_HANDSHAKE(x1)
> DO_COMP_HANDSHAKE(2)

I'd rather not have gotos inside macros, I find that very confusing.
I can add the ifs I suggested if you want the array removed.

Re: [PATCH for-next v3 3/4] RDMA/efa: User/kernel compatibility handshake mechanism

[Jason Gunthorpe]

On Wed, Jul 22, 2020 at 03:13:57PM +0300, Gal Pressman wrote:
> On 22/07/2020 15:08, Jason Gunthorpe wrote:
> > On Wed, Jul 22, 2020 at 03:04:20PM +0300, Gal Pressman wrote:
> >> On 22/07/2020 14:55, Jason Gunthorpe wrote:
> >>> On Tue, Jul 21, 2020 at 04:30:48PM +0300, Gal Pressman wrote:
> >>>> Introduce a mechanism that performs an handshake between the userspace
> >>>> provider and kernel driver which verifies that the user supports all
> >>>> required features in order to operate correctly.
> >>>>
> >>>> The handshake verifies the needed functionality by comparing the
> >>>> reported device caps and the provider caps. If the device reports a
> >>>> non-zero capability the appropriate comp mask is required from the
> >>>> userspace provider in order to allocate the context.
> >>>>
> >>>> Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
> >>>> Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
> >>>> Signed-off-by: Gal Pressman <galpress@amazon.com>
> >>>>  drivers/infiniband/hw/efa/efa_verbs.c | 49 +++++++++++++++++++++++++++
> >>>>  include/uapi/rdma/efa-abi.h           | 10 ++++++
> >>>>  2 files changed, 59 insertions(+)
> >>>>
> >>>> diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
> >>>> index 26102ab333b2..7ca40df81ee5 100644
> >>>> +++ b/drivers/infiniband/hw/efa/efa_verbs.c
> >>>> @@ -1501,11 +1501,48 @@ static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
> >>>>  	return efa_com_dealloc_uar(&dev->edev, &params);
> >>>>  }
> >>>>  
> >>>> +#define EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)                         \
> >>>> +	(!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask)))
> >>>> +
> >>>> +#define DEFINE_COMP_HANDSHAKE(_dev, _comp_mask, _attr, _mask)                  \
> >>>> +	{                                                                      \
> >>>> +		.attr = #_attr,                                                \
> >>>> +		.check_comp = EFA_CHECK_COMP(_dev, _comp_mask, _attr, _mask)   \
> >>>> +	}
> >>>> +
> >>>> +int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
> >>>> +			    const struct efa_ibv_alloc_ucontext_cmd *cmd)
> >>>> +{
> >>>> +	struct efa_dev *dev = to_edev(ibucontext->device);
> >>>> +	int i;
> >>>> +	struct {
> >>>> +		char *attr;
> >>>> +		bool check_comp;
> >>>> +	} user_comp_handshakes[] = {
> >>>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, max_tx_batch,
> >>>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH),
> >>>> +		DEFINE_COMP_HANDSHAKE(dev, cmd->comp_mask, min_sq_depth,
> >>>> +				      EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR),
> >>>> +	};
> >>>
> >>> This seems like a very expensive construct
> >>>
> >>> Why have the array at all? Just list the macros and have them jump to
> >>> err
> >>
> >> Do you mean:
> >>
> >> if (CHECK_COMP(x1)) {
> >>     ibdev_dbg(err);
> >>     goto err;
> >> }
> >>
> >> if (CHECK_COMP(x2)) {
> >>     ibdev_dbg(err);
> >>     goto err;
> >> }
> >>
> >> [...]
> >>
> >> That adds much more boilerplate code for each feature. Or do you have something
> >> else in mind?
> > 
> > #define DO_COMP_HANDSHAKE() \
> >     if (...) goto err
> > 
> > DO_COMP_HANDSHAKE(x1)
> > DO_COMP_HANDSHAKE(2)
> 
> I'd rather not have gotos inside macros, I find that very confusing.
> I can add the ifs I suggested if you want the array removed.

IHMO it barely seems worth all the macros in the first place, and
constructing that whole array on the stack at runtime is pretty
horrible

Jason

[PATCH for-next v3 4/4] RDMA/efa: Add EFA 0xefa1 PCI ID

[Gal Pressman]

Add support for 0xefa1 devices.

Reviewed-by: Shadi Ammouri <sammouri@amazon.com>
Reviewed-by: Yossi Leybovich <sleybo@amazon.com>
Signed-off-by: Gal Pressman <galpress@amazon.com>
---
 drivers/infiniband/hw/efa/efa_main.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c
index 82145574c928..92d701146320 100644
--- a/drivers/infiniband/hw/efa/efa_main.c
+++ b/drivers/infiniband/hw/efa/efa_main.c
@@ -12,10 +12,12 @@
 
 #include "efa.h"
 
-#define PCI_DEV_ID_EFA_VF 0xefa0
+#define PCI_DEV_ID_EFA0_VF 0xefa0
+#define PCI_DEV_ID_EFA1_VF 0xefa1
 
 static const struct pci_device_id efa_pci_tbl[] = {
-	{ PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA_VF) },
+	{ PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA0_VF) },
+	{ PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA1_VF) },
 	{ }
 };
 
-- 
2.27.0