[dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

[tone.zhang]

With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to
the PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.

This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.

The patch fixes the debug log as below:
EAL: Skipping BAR0

Signed-off-by: tone.zhang <***@arm.com>
Reviewed-by: Gavin Hu <***@arm.com>
Reviewed-by: Honnappa Nagarahalli <***@arm.com>
Reviewed-by: Steve Capper <***@arm.com>
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)

diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];

- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;
+ }

if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +476,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,

memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }

RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

[Burakov, Anatoly]

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to
the PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be
enough :)

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be
unaligned, so falling back to this value may cause mapping failure,
because we later use this value as a size of mapping (which needs to be
page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?

Post by tone.zhang
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +476,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

[Tone Zhang (Arm Technology China)]

Hi Burakov,

I'm sorry for the late response.

Thanks a lot for your comments. Please find my response below (marked with "Tone:"). 😊

Br,
Tone

-----Original Message-----
From: Burakov, Anatoly <***@intel.com>
Sent: Wednesday, October 24, 2018 5:09 PM
To: Tone Zhang (Arm Technology China) <***@arm.com>; ***@dpdk.org
Cc: nd <***@arm.com>
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

With a larger PAGE_SIZE it is possible for the MSI table to very close
to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the end
offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)

Tone: Will update code and remove "this" in next version.

+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.

Tone: Sure, it is better! Will update code in next version. Thanks!

+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?

Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +476,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

--

[Burakov, Anatoly]

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below (marked with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

With a larger PAGE_SIZE it is possible for the MSI table to very close
to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the end
offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).

However, i think it's still worth aligning (and putting in an additional
check), because we want to make sure we *don't* attempt to map the MSI-X
BAR, and kernel might do that by adjusting length automatically and
return mmap failure that way.

[Tone Zhang (Arm Technology China)]

Hi Burakov,

Thanks!
Please check my feedback below.

Br,
Tone

-----Original Message-----
From: dev <dev-***@dpdk.org> On Behalf Of Burakov, Anatoly
Sent: Thursday, November 1, 2018 6:01 PM
To: Tone Zhang (Arm Technology China) <***@arm.com>; ***@dpdk.org
Cc: nd <***@arm.com>
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below (marked with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).

However, i think it's still worth aligning (and putting in an additional check), because we want to make sure we *don't* attempt to map the MSI-X BAR, and kernel might do that by adjusting length automatically and return mmap failure that way.

Tone: Thanks a lot! I agree with you. It worth aligning the size. I will update code (RTE_ALIGN_FLOOR by page size) in next version.
I'd like to discuss one case with you. In the case, base->size is 16384, msix_table->offset is 8192, page_size is 65536. After align "msix_table->offset" with page_size (RTE_ALIGN_FLOOR), the value of "table_start" is 0, mmap() will report error, and the memory mapping is failed.
For the case (table_start is 0 after the aglinment), may I continue falling back the "table_start" to " msix_table->offset" (not aligned with page size), and left system adjust the leng

[Burakov, Anatoly]

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Thanks!
Please check my feedback below.
Br,
Tone
-----Original Message-----
Sent: Thursday, November 1, 2018 6:01 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below (marked with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).
However, i think it's still worth aligning (and putting in an additional check), because we want to make sure we *don't* attempt to map the MSI-X BAR, and kernel might do that by adjusting length automatically and return mmap failure that way.
Tone: Thanks a lot! I agree with you. It worth aligning the size. I will update code (RTE_ALIGN_FLOOR by page size) in next version.
I'd like to discuss one case with you. In the case, base->size is 16384, msix_table->offset is 8192, page_size is 65536. After align "msix_table->offset" with page_size (RTE_ALIGN_FLOOR), the value of "table_start" is 0, mmap() will report error, and the memory mapping is failed.
For the case (table_start is 0 after the aglinment), may I continue falling back the "table_start" to " msix_table->offset" (not aligned with page size), and left system adjust the length automatically? Thanks!

Please correct me if i'm wrong, but this is a code path for when we're
trying to mmap around the MSI-X BAR. Kernel will not allow us to do
that, period, so whatever start/end addresses you get, they *must not*
include a single byte of MSI-X BAR. So, in case like you described, i
think we should just straight up refuse the map the entire BAR.

However, as i do not have a system with such properties to test on, so
please correct me if i'm wrong here :)

Post by Tone Zhang (Arm Technology China)
--
Thanks,
Anatoly

[Tone Zhang (Arm Technology China)]

Hi Burakov,

Please find my test case below. Thanks!

Br,
Tone

-----Original Message-----
From: Burakov, Anatoly <***@intel.com>
Sent: Tuesday, November 6, 2018 7:03 PM
To: Tone Zhang (Arm Technology China) <***@arm.com>; ***@dpdk.org
Cc: nd <***@arm.com>
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Thanks!
Please check my feedback below.
Br,
Tone
-----Original Message-----
Sent: Thursday, November 1, 2018 6:01 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below (marked with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).
However, i think it's still worth aligning (and putting in an additional check), because we want to make sure we *don't* attempt to map the MSI-X BAR, and kernel might do that by adjusting length automatically and return mmap failure that way.
Tone: Thanks a lot! I agree with you. It worth aligning the size. I will update code (RTE_ALIGN_FLOOR by page size) in next version.
I'd like to discuss one case with you. In the case, base->size is 16384, msix_table->offset is 8192, page_size is 65536. After align "msix_table->offset" with page_size (RTE_ALIGN_FLOOR), the value of "table_start" is 0, mmap() will report error, and the memory mapping is failed.
For the case (table_start is 0 after the aglinment), may I continue falling back the "table_start" to " msix_table->offset" (not aligned with page size), and left system adjust the length automatically? Thanks!

Please correct me if i'm wrong, but this is a code path for when we're trying to mmap around the MSI-X BAR. Kernel will not allow us to do that, period, so whatever start/end addresses you get, they *must not* include a single byte of MSI-X BAR. So, in case like you described, i think we should just straight up refuse the map the entire BAR.

However, as i do not have a system with such properties to test on, so please correct me if i'm wrong here :)


Tone: I understand and agree with you. 😊

Please have a look at my test case. In my case, I tried to bind NVMe device with VFIO driver and the kernel page size is 64KB. Without the change, the test is failed.

From the debug information, I observed that "bar->size" is 16384, "msix_table->offset" is 8192 and "msix_table->size" is 512. Regarding the page size is much bigger than the "bar->size", in the change, the code maps the first 8192 bytes ahead of MSI-X table. After align with the page size boundary, the "start" offset after the MSI-X table is over "bar->size", mmap() reports error. In this case, I can only map the memory before the MSI-X table. After fall back "table_start" to " msix_table->offset " (i.e. 8192 bytes), and NOT mapping the memory behind MSI-X table, the NVMe device can be bound to VFIO driver, and the test is passed. The kernel version in my test environment is 4.16.

So in the change, I do not map any byte of MSI-X table, unfortunately I cannot align the memory "size" in mmap() to page size boundary. From the test result, the change fixes the error. The case looks a little tricky. If we refuse the memory map here, it means we cannot bind VFIO driver with some PCI devices with 64KB kernel page size. I hope we can support such case in DPDK. 😊

Post by Tone Zhang (Arm Technology China)
--
Thank

[Burakov, Anatoly]

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Please find my test case below. Thanks!
Br,
Tone
-----Original Message-----
Sent: Tuesday, November 6, 2018 7:03 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Thanks!
Please check my feedback below.
Br,
Tone
-----Original Message-----
Sent: Thursday, November 1, 2018 6:01 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below (marked
with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).
However, i think it's still worth aligning (and putting in an additional check), because we want to make sure we *don't* attempt to map the MSI-X BAR, and kernel might do that by adjusting length automatically and return mmap failure that way.
Tone: Thanks a lot! I agree with you. It worth aligning the size. I will update code (RTE_ALIGN_FLOOR by page size) in next version.
I'd like to discuss one case with you. In the case, base->size is 16384, msix_table->offset is 8192, page_size is 65536. After align "msix_table->offset" with page_size (RTE_ALIGN_FLOOR), the value of "table_start" is 0, mmap() will report error, and the memory mapping is failed.
For the case (table_start is 0 after the aglinment), may I continue falling back the "table_start" to " msix_table->offset" (not aligned with page size), and left system adjust the length automatically? Thanks!

Please correct me if i'm wrong, but this is a code path for when we're trying to mmap around the MSI-X BAR. Kernel will not allow us to do that, period, so whatever start/end addresses you get, they *must not* include a single byte of MSI-X BAR. So, in case like you described, i think we should just straight up refuse the map the entire BAR.
However, as i do not have a system with such properties to test on, so please correct me if i'm wrong here :)
Tone: I understand and agree with you. 😊
Please have a look at my test case. In my case, I tried to bind NVMe device with VFIO driver and the kernel page size is 64KB. Without the change, the test is failed.
From the debug information, I observed that "bar->size" is 16384, "msix_table->offset" is 8192 and "msix_table->size" is 512. Regarding the page size is much bigger than the "bar->size", in the change, the code maps the first 8192 bytes ahead of MSI-X table. After align with the page size boundary, the "start" offset after the MSI-X table is over "bar->size", mmap() reports error. In this case, I can only map the memory before the MSI-X table. After fall back "table_start" to " msix_table->offset " (i.e. 8192 bytes), and NOT mapping the memory behind MSI-X table, the NVMe device can be bound to VFIO driver, and the test is passed. The kernel version in my test environment is 4.16.
So in the change, I do not map any byte of MSI-X table, unfortunately I cannot align the memory "size" in mmap() to page size boundary. From the test result, the change fixes the error. The case looks a little tricky. If we refuse the memory map here, it means we cannot bind VFIO driver with some PCI devices with 64KB kernel page size. I hope we can support such case in DPDK. 😊

Hi Tone,

If it works and doesn't impact any other cases, i'm happy to include the
above change (i.e. fall back to unaligned offset if aligning it results
in zero offset). I'm curious what would happen if there's something else
after the MSI-X table as well, and what alignments would be required for
that... But i would rather wait for someone to come to us with an actual
test case for that as well :)

Post by Tone Zhang (Arm Technology China)

Post by Tone Zhang (Arm Technology China)
--
Thanks,
Anatoly

--
Thanks,
Anatoly

[Tone Zhang (Arm Technology China)]

Hi Burakov,

Thanks a lot for you review. I will update the change and push the next version ASAP.

Thanks!

Br,
Tone

-----Original Message-----
From: Burakov, Anatoly <***@intel.com>
Sent: Wednesday, November 7, 2018 6:13 PM
To: Tone Zhang (Arm Technology China) <***@arm.com>; ***@dpdk.org
Cc: nd <***@arm.com>
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Please find my test case below. Thanks!
Br,
Tone
-----Original Message-----
Sent: Tuesday, November 6, 2018 7:03 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
Thanks!
Please check my feedback below.
Br,
Tone
-----Original Message-----
Sent: Thursday, November 1, 2018 6:01 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by Tone Zhang (Arm Technology China)
Hi Burakov,
I'm sorry for the late response.
Thanks a lot for your comments. Please find my response below
(marked with "Tone:"). 😊
Br,
Tone
-----Original Message-----
Sent: Wednesday, October 24, 2018 5:09 PM
Subject: Re: [dpdk-dev] [PATCH] pci_vfio: Support 64KB kernel
page_size with vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out the PCI BAR
boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 25 +++++++++++++++++++++----
1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index b1f0683..1373345 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping this BAR%d\n", bar_index);
return 0;

I feel like "this" is unnecessary here - just "Skipping BAR%d" should be enough :)
Tone: Will update code and remove "this" in next version.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -457,7 +459,12 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_start = (table_start + ~PAGE_MASK) & PAGE_MASK;

IMO these two additions should be replaced by RTE_ALIGN by page size.
Makes the purpose of the code much clearer.
Tone: Sure, it is better! Will update code in next version. Thanks!

Post by tone.zhang
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

If i understand things correctly, msix_table->offset value here may be unaligned, so falling back to this value may cause mapping failure, because we later use this value as a size of mapping (which needs to be page aligned). Shouldn't this be aligned using RTE_ALIGN_FLOOR by page size?
Tone: It is a little tricky. Align msix_table->offset with RTE_ALIGN_FLOOR maybe get 0 if the offset is less than page size in the PCI bar. It will trigger mmap() error. IIRC the input parameter "size" in mmap() is not required to be aligned with page size, system will do it. But it is better if we can do it. If I was wrong, please correct me. Thanks a lot.

Apologies, you're correct - length can be misaligned (just tested it).
However, i think it's still worth aligning (and putting in an additional check), because we want to make sure we *don't* attempt to map the MSI-X BAR, and kernel might do that by adjusting length automatically and return mmap failure that way.
Tone: Thanks a lot! I agree with you. It worth aligning the size. I will update code (RTE_ALIGN_FLOOR by page size) in next version.
I'd like to discuss one case with you. In the case, base->size is 16384, msix_table->offset is 8192, page_size is 65536. After align "msix_table->offset" with page_size (RTE_ALIGN_FLOOR), the value of "table_start" is 0, mmap() will report error, and the memory mapping is failed.
For the case (table_start is 0 after the aglinment), may I continue falling back the "table_start" to " msix_table->offset" (not aligned with page size), and left system adjust the length automatically? Thanks!

Please correct me if i'm wrong, but this is a code path for when we're trying to mmap around the MSI-X BAR. Kernel will not allow us to do that, period, so whatever start/end addresses you get, they *must not* include a single byte of MSI-X BAR. So, in case like you described, i think we should just straight up refuse the map the entire BAR.
However, as i do not have a system with such properties to test on, so
please correct me if i'm wrong here :)
Tone: I understand and agree with you. 😊
Please have a look at my test case. In my case, I tried to bind NVMe device with VFIO driver and the kernel page size is 64KB. Without the change, the test is failed.
From the debug information, I observed that "bar->size" is 16384, "msix_table->offset" is 8192 and "msix_table->size" is 512. Regarding the page size is much bigger than the "bar->size", in the change, the code maps the first 8192 bytes ahead of MSI-X table. After align with the page size boundary, the "start" offset after the MSI-X table is over "bar->size", mmap() reports error. In this case, I can only map the memory before the MSI-X table. After fall back "table_start" to " msix_table->offset " (i.e. 8192 bytes), and NOT mapping the memory behind MSI-X table, the NVMe device can be bound to VFIO driver, and the test is passed. The kernel version in my test environment is 4.16.
So in the change, I do not map any byte of MSI-X table, unfortunately
I cannot align the memory "size" in mmap() to page size boundary. From
the test result, the change fixes the error. The case looks a little
tricky. If we refuse the memory map here, it means we cannot bind VFIO
driver with some PCI devices with 64KB kernel page size. I hope we can
support such case in DPDK. 😊

Hi Tone,

If it works and doesn't impact any other cases, i'm happy to include the above change (i.e. fall back to unaligned offset if aligning it results in zero offset). I'm curious what would happen if there's something else after the MSI-X table as well, and what alignments would be required for that... But i would rather wait for someone to come to us with an actual test case for that as well :)

Post by Tone Zhang (Arm Technology China)

Post by Tone Zhang (Arm Technology China)
--
Thanks,
Anatoly

--
Thanks,
Anatoly

[tone.zhang]

With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end
of the MSI table to the PAGE_SIZE, the end offset of the MSI
table is out of the PCI BAR boundary.

This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.

The patch fixes the debug log as below:
EAL: Skipping BAR0

Signed-off-by: tone.zhang <***@arm.com>
Reviewed-by: Gavin Hu <***@arm.com>
Reviewed-by: Honnappa Nagarahalli <***@arm.com>
Reviewed-by: Steve Capper <***@arm.com>
Reviewed-by: Burakov Anatoly <***@intel.com>
---
drivers/bus/pci/linux/pci_vfio.c | 36 +++++++++++++++++++++++++++++++-----
1 file changed, 31 insertions(+), 5 deletions(-)

diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index 305cc06..9a0affe 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];

- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }

if (msix_table->bar_index == bar_index) {
/*
@@ -456,8 +458,22 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar and
+ * align with PAGE_SIZE.
+ */
+ if (table_start >= bar->size) {
+ table_start = RTE_ALIGN_FLOOR(msix_table->offset,
+ PAGE_SIZE);
+ /* after aligning with PAGE_SIZE, if it is less than
+ * the MSI-X table offset, continue falling back to
+ * the actual MSI-X table offset in the bar.
+ */
+ if (table_start < msix_table->offset)
+ table_start = msix_table->offset;
+ }

if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +485,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,

memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }

RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

[Burakov, Anatoly]

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end
of the MSI table to the PAGE_SIZE, the end offset of the MSI
table is out of the PCI BAR boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.
EAL: Skipping BAR0

In the future, please don't include my Reviewed tag unless i actually
sent one :)

Post by tone.zhang
---
drivers/bus/pci/linux/pci_vfio.c | 36 +++++++++++++++++++++++++++++++-----
1 file changed, 31 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index 305cc06..9a0affe 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -456,8 +458,22 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar and
+ * align with PAGE_SIZE.
+ */

Minor nitpick - wording of comment could be better, for example:

if page-aligned start of MSI-X table is beyond BAR size, shrink the
mapping size to MSI-X table start address.

Also, probably needs newline before comment.

Post by tone.zhang
+ if (table_start >= bar->size) {
+ table_start = RTE_ALIGN_FLOOR(msix_table->offset,
+ PAGE_SIZE);
+ /* after aligning with PAGE_SIZE, if it is less than
+ * the MSI-X table offset, continue falling back to
+ * the actual MSI-X table offset in the bar.
+ */

Same here, wording could probably be improved. Suggested rewording:

If MSI-X table address, floor-aligned by page size, is lower than actual
MSI-X table offset, fall back to using MSI-X table offset as table start.

Now that i think of it, this could really be expressed like this:

uint32_t aligned = RTE_ALIGN_FLOOR(msix_table->offset, PAGE_SIZE);
table_start = RTE_MAX(aligned, msix_table_offset);

I believe this would be much clearer.

Post by tone.zhang
+ if (table_start < msix_table->offset)
+ table_start = msix_table->offset;
+ }
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +485,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */

Suggested rewording:

If MSI-X table end is beyond BAR end, don't attempt to perform second
mapping.

Post by tone.zhang
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

However, the patch can go in as is if needed, so

Reviewed-by: Anatoly Burakov <***@intel.com>

[Tone Zhang (Arm Technology China)]

Hi Anatoly,

Sorry for the late response.

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
Sent: Friday, November 9, 2018 8:15 PM
Subject: Re: [PATCH v2] pci_vfio: Support 64KB kernel page_size with vfio-pci
driver

With a larger PAGE_SIZE it is possible for the MSI table to very close
to the end of the BAR s.t. when we align the start and end of the MSI
table to the PAGE_SIZE, the end offset of the MSI table is out of the
PCI BAR boundary.
This patch addresses the issue by comparing both the start and the end
offset of the MSI table with the BAR size, and skip the mapping if it
is out of Bar scope.
EAL: Skipping BAR0

In the future, please don't include my Reviewed tag unless i actually sent one :)

Thanks a lot! Will keep in mind. 😊

Post by Tone Zhang (Arm Technology China)

---
drivers/bus/pci/linux/pci_vfio.c | 36 +++++++++++++++++++++++++++++++---

--

1 file changed, 31 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index 305cc06..9a0affe 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -456,8 +458,22 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar and
+ * align with PAGE_SIZE.
+ */

if page-aligned start of MSI-X table is beyond BAR size, shrink the mapping size
to MSI-X table start address.
Also, probably needs newline before comment.

Will update the code in next version. Thanks!

Post by Tone Zhang (Arm Technology China)

+ if (table_start >= bar->size) {
+ table_start = RTE_ALIGN_FLOOR(msix_table->offset,
+ PAGE_SIZE);
+ /* after aligning with PAGE_SIZE, if it is less than
+ * the MSI-X table offset, continue falling back to
+ * the actual MSI-X table offset in the bar.
+ */

If MSI-X table address, floor-aligned by page size, is lower than actual MSI-X
table offset, fall back to using MSI-X table offset as table start.
uint32_t aligned = RTE_ALIGN_FLOOR(msix_table->offset, PAGE_SIZE);
table_start = RTE_MAX(aligned, msix_table_offset);
I believe this would be much clearer.

Will update the patch.

Post by Tone Zhang (Arm Technology China)

+ if (table_start < msix_table->offset)
+ table_start = msix_table->offset;
+ }
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +485,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */

If MSI-X table end is beyond BAR end, don't attempt to perform second mapping.

Thanks a lot. Will update.

Post by Tone Zhang (Arm Technology China)

+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

However, the patch can go in as is if needed, so

Thanks! 😉

Post by Tone Zhang (Arm Technology China)
--
Thanks,
Anatoly

Br,

[Tone Zhang (Arm Technology China)]

Hi Anatoly,

I have some comments.

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
From: Tone Zhang (Arm Technology China)
Sent: Thursday, November 15, 2018 8:49 AM
Subject: RE: [PATCH v2] pci_vfio: Support 64KB kernel page_size with vfio-pci
driver
Hi Anatoly,
Sorry for the late response.

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
Sent: Friday, November 9, 2018 8:15 PM
Subject: Re: [PATCH v2] pci_vfio: Support 64KB kernel page_size with
vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end of
the MSI table to the PAGE_SIZE, the end offset of the MSI table is
out of the PCI BAR boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.
EAL: Skipping BAR0

In the future, please don't include my Reviewed tag unless i actually sent one :)

Thanks a lot! Will keep in mind. 😊

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
---
drivers/bus/pci/linux/pci_vfio.c | 36
+++++++++++++++++++++++++++++++---

--

Post by tone.zhang
1 file changed, 31 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index 305cc06..9a0affe 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

Post by tone.zhang
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -456,8 +458,22 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

Post by tone.zhang
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar and
+ * align with PAGE_SIZE.
+ */

if page-aligned start of MSI-X table is beyond BAR size, shrink the
mapping size to MSI-X table start address.
Also, probably needs newline before comment.

Will update the code in next version. Thanks!

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ if (table_start >= bar->size) {
+ table_start = RTE_ALIGN_FLOOR(msix_table->offset,
+ PAGE_SIZE);
+ /* after aligning with PAGE_SIZE, if it is less than
+ * the MSI-X table offset, continue falling back to
+ * the actual MSI-X table offset in the bar.
+ */

If MSI-X table address, floor-aligned by page size, is lower than
actual MSI-X table offset, fall back to using MSI-X table offset as table start.
uint32_t aligned = RTE_ALIGN_FLOOR(msix_table->offset, PAGE_SIZE);
table_start = RTE_MAX(aligned, msix_table_offset);
I believe this would be much clearer.

Will update the patch.

When enter the judgement, it implies the "msix_table->offset" is NOT page size aligned, I think we can replace the code in the judgement with one line: table_start = msix_table->offset;
It looks more simple. What's your opinion? Thanks!

Post by Tone Zhang (Arm Technology China)

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ if (table_start < msix_table->offset)
+ table_start = msix_table->offset;
+ }
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +485,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */

If MSI-X table end is beyond BAR end, don't attempt to perform second

mapping.
Thanks a lot. Will update.

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

However, the patch can go in as is if needed, so

Thanks! 😉

Post by Tone Zhang (Arm Technology China)
--
Thanks,

[Burakov, Anatoly]

Post by Tone Zhang (Arm Technology China)
Hi Anatoly,
I have some comments.

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
From: Tone Zhang (Arm Technology China)
Sent: Thursday, November 15, 2018 8:49 AM
Subject: RE: [PATCH v2] pci_vfio: Support 64KB kernel page_size with vfio-pci
driver
Hi Anatoly,
Sorry for the late response.

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
Sent: Friday, November 9, 2018 8:15 PM
Subject: Re: [PATCH v2] pci_vfio: Support 64KB kernel page_size with
vfio-pci driver

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end of
the MSI table to the PAGE_SIZE, the end offset of the MSI table is
out of the PCI BAR boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.
EAL: Skipping BAR0

In the future, please don't include my Reviewed tag unless i actually sent one :)

Thanks a lot! Will keep in mind. 😊

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
---
drivers/bus/pci/linux/pci_vfio.c | 36
+++++++++++++++++++++++++++++++---

--

Post by tone.zhang
1 file changed, 31 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c
b/drivers/bus/pci/linux/pci_vfio.c
index 305cc06..9a0affe 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -445,9 +445,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

Post by tone.zhang
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -456,8 +458,22 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct

mapped_pci_resource *vfio_res,

Post by tone.zhang
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+ /* after rounding to PAGE_SIZE, it is over the bar->size,
+ * fall back to the MSI-X table offset in the bar and
+ * align with PAGE_SIZE.
+ */

if page-aligned start of MSI-X table is beyond BAR size, shrink the
mapping size to MSI-X table start address.
Also, probably needs newline before comment.

Will update the code in next version. Thanks!

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ if (table_start >= bar->size) {
+ table_start = RTE_ALIGN_FLOOR(msix_table->offset,
+ PAGE_SIZE);
+ /* after aligning with PAGE_SIZE, if it is less than
+ * the MSI-X table offset, continue falling back to
+ * the actual MSI-X table offset in the bar.
+ */

If MSI-X table address, floor-aligned by page size, is lower than
actual MSI-X table offset, fall back to using MSI-X table offset as table start.
uint32_t aligned = RTE_ALIGN_FLOOR(msix_table->offset, PAGE_SIZE);
table_start = RTE_MAX(aligned, msix_table_offset);
I believe this would be much clearer.

Will update the patch.

When enter the judgement, it implies the "msix_table->offset" is NOT page size aligned, I think we can replace the code in the judgement with one line: table_start = msix_table->offset;
It looks more simple. What's your opinion? Thanks!

Agree, what was i thinking :D

Post by Tone Zhang (Arm Technology China)

Post by Tone Zhang (Arm Technology China)

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ if (table_start < msix_table->offset)
+ table_start = msix_table->offset;
+ }
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -469,8 +485,18 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * after rounding to PAGE_SIZE we don't have any space
+ * left after the MSI table, so don't try and map it.
+ */

If MSI-X table end is beyond BAR end, don't attempt to perform second

mapping.
Thanks a lot. Will update.

Post by Tone Zhang (Arm Technology China)

Post by tone.zhang
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ }
+ else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

However, the patch can go in as is if needed, so

Thanks! 😉

Post by Tone Zhang (Arm Technology China)
--
Thanks,
Anatoly

Br,
Tone

Br,
Tone

[tone.zhang]

With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end
of the MSI table to the PAGE_SIZE, the end offset of the MSI
table is out of the PCI BAR boundary.

This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.

The patch fixes the debug log as below:
EAL: Skipping BAR0

Signed-off-by: tone.zhang <***@arm.com>
Reviewed-by: Gavin Hu <***@arm.com>
Reviewed-by: Honnappa Nagarahalli <***@arm.com>
Reviewed-by: Steve Capper <***@arm.com>
Reviewed-by: Anatoly Burakov <***@intel.com>
---
drivers/bus/pci/linux/pci_vfio.c | 27 ++++++++++++++++++++++-----
1 file changed, 22 insertions(+), 5 deletions(-)

diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index ffd26f1..0e7bfd7 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -457,9 +457,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];

- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }

if (msix_table->bar_index == bar_index) {
/*
@@ -468,8 +470,14 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+
+ /* If page-aligned start of MSI-X table is beyond BAR size,
+ * shrink the mapping size to MSI-X table start address.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;

if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -481,8 +489,17 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,

memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * If MSI-X table end is beyond BAR end, don't attempt
+ * to perform second mapping.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ } else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }

RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

[Tone Zhang (Arm Technology China)]

Hi all,

Could anyone please have a review and re-test the change?

Thanks a lot!

Br,
Tone

Post by Tone Zhang (Arm Technology China)
-----Original Message-----
Sent: Monday, November 19, 2018 10:37 AM
Subject: [dpdk-dev] [PATCH v3] pci_vfio: Support 64KB kernel page_size with
vfio-pci driver
With a larger PAGE_SIZE it is possible for the MSI table to very close to the end
of the BAR s.t. when we align the start and end of the MSI table to the
PAGE_SIZE, the end offset of the MSI table is out of the PCI BAR boundary.
This patch addresses the issue by comparing both the start and the end offset of
the MSI table with the BAR size, and skip the mapping if it is out of Bar scope.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 27 ++++++++++++++++++++++-----
1 file changed, 22 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index ffd26f1..0e7bfd7 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -457,9 +457,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -468,8 +470,14 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);
+
+ /* If page-aligned start of MSI-X table is beyond BAR size,
+ * shrink the mapping size to MSI-X table start address.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -481,8 +489,17 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct
mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * If MSI-X table end is beyond BAR end, don't attempt
+ * to perform second mapping.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ } else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "
--
2.7.4

[Burakov, Anatoly]

Post by Tone Zhang (Arm Technology China)
Hi all,
Could anyone please have a review and re-test the change?
Thanks a lot!
Br,
Tone

Tested with latest 18.11, all seems to work fine.

Tested-by: Anatoly Burakov <***@intel.com>

[Burakov, Anatoly]

Post by tone.zhang
With a larger PAGE_SIZE it is possible for the MSI table to very
close to the end of the BAR s.t. when we align the start and end
of the MSI table to the PAGE_SIZE, the end offset of the MSI
table is out of the PCI BAR boundary.
This patch addresses the issue by comparing both the start and the
end offset of the MSI table with the BAR size, and skip the mapping
if it is out of Bar scope.
EAL: Skipping BAR0
---
drivers/bus/pci/linux/pci_vfio.c | 27 ++++++++++++++++++++++-----
1 file changed, 22 insertions(+), 5 deletions(-)
diff --git a/drivers/bus/pci/linux/pci_vfio.c b/drivers/bus/pci/linux/pci_vfio.c
index ffd26f1..0e7bfd7 100644
--- a/drivers/bus/pci/linux/pci_vfio.c
+++ b/drivers/bus/pci/linux/pci_vfio.c
@@ -457,9 +457,11 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
struct pci_msix_table *msix_table = &vfio_res->msix_table;
struct pci_map *bar = &vfio_res->maps[bar_index];
- if (bar->size == 0)
+ if (bar->size == 0) {
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;

I would perhaps make it a DEBUG rather than INFO.

Post by tone.zhang
+ }
if (msix_table->bar_index == bar_index) {
/*
@@ -468,8 +470,14 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
*/
uint32_t table_start = msix_table->offset;
uint32_t table_end = table_start + msix_table->size;
- table_end = (table_end + ~PAGE_MASK) & PAGE_MASK;
- table_start &= PAGE_MASK;
+ table_end = RTE_ALIGN(table_end, PAGE_SIZE);
+ table_start = RTE_ALIGN(table_start, PAGE_SIZE);

I believe table_start &= PAGE_MASK is equivalent to table_start =
RTE_ALIGN_FLOOR(), not RTE_ALIGN() (which is an alias for RTE_ALIGN_CEIL()).

Post by tone.zhang
+
+ /* If page-aligned start of MSI-X table is beyond BAR size,
+ * shrink the mapping size to MSI-X table start address.
+ */
+ if (table_start >= bar->size)
+ table_start = msix_table->offset;
if (table_start == 0 && table_end >= bar->size) {
/* Cannot map this BAR */
@@ -481,8 +489,17 @@ pci_vfio_mmap_bar(int vfio_dev_fd, struct mapped_pci_resource *vfio_res,
memreg[0].offset = bar->offset;
memreg[0].size = table_start;
- memreg[1].offset = bar->offset + table_end;
- memreg[1].size = bar->size - table_end;
+ if (bar->size < table_end) {
+ /*
+ * If MSI-X table end is beyond BAR end, don't attempt
+ * to perform second mapping.
+ */
+ memreg[1].offset = 0;
+ memreg[1].size = 0;
+ } else {
+ memreg[1].offset = bar->offset + table_end;
+ memreg[1].size = bar->size - table_end;
+ }
RTE_LOG(DEBUG, EAL,
"Trying to map BAR%d that contains the MSI-X "

[Stephen Hemminger]

On Mon, 10 Dec 2018 11:45:37 +0000

Post by Burakov, Anatoly

Post by tone.zhang
/* Skip this BAR */
+ RTE_LOG(INFO, EAL, "Skipping BAR%d\n", bar_index);
return 0;

I would perhaps make it a DEBUG rather than INFO.

And drop the comment...