Re: [PATCH v3 net-next] net: mana: Add a driver for Microsoft Azure Network Adapter (MANA)

From: Dexuan Cui <decui@microsoft.com>
Date: Thu,  8 Apr 2021 15:58:40 -0700

+struct gdma_msg_hdr {
+	u32 hdr_type;
+	u32 msg_type;
+	u16 msg_version;
+	u16 hwc_msg_id;
+	u32 msg_size;
+} __packed;
+
+struct gdma_dev_id {
+	union {
+		struct {
+			u16 type;
+			u16 instance;
+		};
+
+		u32 as_uint32;
+	};
+} __packed;

Please don't  use __packed unless absolutely necessary.  It generates suboptimal code (byte at a time
accesses etc.) and for many of these you don't even need it.

Thank you.

From: David Miller <davem@davemloft.net>
Sent: Thursday, April 8, 2021 4:46 PM
...

quoted

+struct gdma_msg_hdr {
+	u32 hdr_type;
+	u32 msg_type;
+	u16 msg_version;
+	u16 hwc_msg_id;
+	u32 msg_size;
+} __packed;
+
+struct gdma_dev_id {
+	union {
+		struct {
+			u16 type;
+			u16 instance;
+		};
+
+		u32 as_uint32;
+	};
+} __packed;

Please don't  use __packed unless absolutely necessary.  It generates
suboptimal code (byte at a time
accesses etc.) and for many of these you don't even need it.

In the driver code, all the structs/unions marked by __packed are used to
talk with the hardware, so I think __packed is necessary here?

Do you think if it's better if we remove all the __packed, and add
static_assert(sizeof(struct XXX) == YYY) instead? e.g.

@@ -105,7 +105,8 @@ struct gdma_msg_hdr {
        u16 msg_version;
        u16 hwc_msg_id;
        u32 msg_size;
-} __packed;
+};
+static_assert(sizeof(struct gdma_msg_hdr) == 16);

 struct gdma_dev_id {
        union {

Now I'm trying to figure out how other NIC drivers define structs/unions.

Thanks,
Dexuan

From: Dexuan Cui <decui@microsoft.com>
Date: Fri, 9 Apr 2021 00:24:51 +0000

quoted

From: David Miller <davem@davemloft.net>
Sent: Thursday, April 8, 2021 4:46 PM
...

quoted

+struct gdma_msg_hdr {
+	u32 hdr_type;
+	u32 msg_type;
+	u16 msg_version;
+	u16 hwc_msg_id;
+	u32 msg_size;
+} __packed;
+
+struct gdma_dev_id {
+	union {
+		struct {
+			u16 type;
+			u16 instance;
+		};
+
+		u32 as_uint32;
+	};
+} __packed;

Please don't  use __packed unless absolutely necessary.  It generates
suboptimal code (byte at a time
accesses etc.) and for many of these you don't even need it.

In the driver code, all the structs/unions marked by __packed are used to
talk with the hardware, so I think __packed is necessary here?

It actually isan't in many cases, check with and without the __packed directive
and see if anything chasnges.

This won't make sure the structure member offsets are what you expect.

I think you'll have to go through the structures one-by-one by hand to
figure out which ones really require the __packed attribute and which do not.

Thank you.

From: David Miller <davem@davemloft.net>
Sent: Thursday, April 8, 2021 5:41 PM

quoted

...
In the driver code, all the structs/unions marked by __packed are used to
talk with the hardware, so I think __packed is necessary here?

It actually isan't in many cases, check with and without the __packed directive
and see if anything chasnges.

Will do.

quoted

Do you think if it's better if we remove all the __packed, and add
static_assert(sizeof(struct XXX) == YYY) instead? e.g.

@@ -105,7 +105,8 @@ struct gdma_msg_hdr {
        u16 msg_version;
        u16 hwc_msg_id;
        u32 msg_size;
-} __packed;
+};
+static_assert(sizeof(struct gdma_msg_hdr) == 16);

This won't make sure the structure member offsets are what you expect.

I think you'll have to go through the structures one-by-one by hand to
figure out which ones really require the __packed attribute and which do not.

Got it. Let me see if I can remove all the __packed.

-----Original Message-----
From: David Miller <davem@davemloft.net>
Sent: Thursday, April 8, 2021 8:41 PM
To: Dexuan Cui <decui@microsoft.com>
Cc: kuba@kernel.org; KY Srinivasan <kys@microsoft.com>; Haiyang Zhang
[off-list ref]; Stephen Hemminger
[off-list ref]; wei.liu@kernel.org; Wei Liu
[off-list ref]; netdev@vger.kernel.org; leon@kernel.org;
andrew@lunn.ch; bernd@petrovitsch.priv.at; rdunlap@infradead.org; linux-
kernel@vger.kernel.org; linux-hyperv@vger.kernel.org
Subject: Re: [PATCH v3 net-next] net: mana: Add a driver for Microsoft Azure
Network Adapter (MANA)

From: Dexuan Cui <decui@microsoft.com>
Date: Fri, 9 Apr 2021 00:24:51 +0000

quoted

quoted

From: David Miller <davem@davemloft.net>
Sent: Thursday, April 8, 2021 4:46 PM
...

quoted

+struct gdma_msg_hdr {
+	u32 hdr_type;
+	u32 msg_type;
+	u16 msg_version;
+	u16 hwc_msg_id;
+	u32 msg_size;
+} __packed;
+
+struct gdma_dev_id {
+	union {
+		struct {
+			u16 type;
+			u16 instance;
+		};
+
+		u32 as_uint32;
+	};
+} __packed;

Please don't  use __packed unless absolutely necessary.  It generates
suboptimal code (byte at a time
accesses etc.) and for many of these you don't even need it.

In the driver code, all the structs/unions marked by __packed are used to
talk with the hardware, so I think __packed is necessary here?

It actually isan't in many cases, check with and without the __packed
directive
and see if anything chasnges.

quoted

Do you think if it's better if we remove all the __packed, and add
static_assert(sizeof(struct XXX) == YYY) instead? e.g.

@@ -105,7 +105,8 @@ struct gdma_msg_hdr {
        u16 msg_version;
        u16 hwc_msg_id;
        u32 msg_size;
-} __packed;
+};
+static_assert(sizeof(struct gdma_msg_hdr) == 16);

This won't make sure the structure member offsets are what you expect.

I think you'll have to go through the structures one-by-one by hand to
figure out which ones really require the __packed attribute and which do not.

For the structs containing variables with the same sizes, or already size aligned 
variables, we knew the __packed has no effect. And for these structs, it doesn't 
cause performance impact either, correct? 

But in the future, if different sized variables are added, the __packed may 
become necessary again. To prevent anyone accidently forget to add __packed 
when adding new variables to these structs, can we keep the __packed for all 
messages going through the "wire"?

Thanks,
- Haiyang

 For the structs containing variables with the same sizes, or already size aligned 

variables, we knew the __packed has no effect. And for these structs, it doesn't 
cause performance impact either, correct? 

But in the future, if different sized variables are added, the __packed may 
become necessary again. To prevent anyone accidently forget to add __packed 
when adding new variables to these structs, can we keep the __packed for all 
messages going through the "wire"?

It should not be a problem because anybody adding new variables should
know packed is not liked in the kernel and will take care.

If you want to be paranoid add a BUILD_BUG_ON(size(struct foo) != 42);

   Andrew