Major changes in v1:
--------------
1. Implemented MVP queue API ndos to remove the userspace-visible
driver reset.
2. Fixed issues in the napi_pp_put_page() devmem frag unref path.
3. Removed RFC tag.
Many smaller addressed comments across all the patches (patches have
individual change log).
Full tree including the rest of the GVE driver changes:
https://github.com/mina/linux/commits/tcpdevmem-v1
Cc: Yunsheng Lin <linyunsheng(a)huawei.com>
Cc: Shailend Chand <shailend(a)google.com>
Cc: Harshitha Ramamurthy <hramamurthy(a)google.com>
Changes in RFC v3:
------------------
1. Pulled in the memory-provider dependency from Jakub's RFC[1] to make the
series reviewable and mergable.
2. Implemented multi-rx-queue binding which was a todo in v2.
3. Fix to cmsg handling.
The sticking point in RFC v2[2] was the device reset required to refill
the device rx-queues after the dmabuf bind/unbind. The solution
suggested as I understand is a subset of the per-queue management ops
Jakub suggested or similar:
https://lore.kernel.org/netdev/20230815171638.4c057dcd@kernel.org/
This is not addressed in this revision, because:
1. This point was discussed at netconf & netdev and there is openness to
using the current approach of requiring a device reset.
2. Implementing individual queue resetting seems to be difficult for my
test bed with GVE. My prototype to test this ran into issues with the
rx-queues not coming back up properly if reset individually. At the
moment I'm unsure if it's a mistake in the POC or a genuine issue in
the virtualization stack behind GVE, which currently doesn't test
individual rx-queue restart.
3. Our usecases are not bothered by requiring a device reset to refill
the buffer queues, and we'd like to support NICs that run into this
limitation with resetting individual queues.
My thought is that drivers that have trouble with per-queue configs can
use the support in this series, while drivers that support new netdev
ops to reset individual queues can automatically reset the queue as
part of the dma-buf bind/unbind.
The same approach with device resets is presented again for consideration
with other sticking points addressed.
This proposal includes the rx devmem path only proposed for merge. For a
snapshot of my entire tree which includes the GVE POC page pool support &
device memory support:
https://github.com/torvalds/linux/compare/master...mina:linux:tcpdevmem-v3
[1] https://lore.kernel.org/netdev/f8270765-a27b-6ccf-33ea-cda097168d79@redhat.…
[2] https://lore.kernel.org/netdev/CAHS8izOVJGJH5WF68OsRWFKJid1_huzzUK+hpKbLcL4…
Cc: Shakeel Butt <shakeelb(a)google.com>
Cc: Jeroen de Borst <jeroendb(a)google.com>
Cc: Praveen Kaligineedi <pkaligineedi(a)google.com>
Changes in RFC v2:
------------------
The sticking point in RFC v1[1] was the dma-buf pages approach we used to
deliver the device memory to the TCP stack. RFC v2 is a proof-of-concept
that attempts to resolve this by implementing scatterlist support in the
networking stack, such that we can import the dma-buf scatterlist
directly. This is the approach proposed at a high level here[2].
Detailed changes:
1. Replaced dma-buf pages approach with importing scatterlist into the
page pool.
2. Replace the dma-buf pages centric API with a netlink API.
3. Removed the TX path implementation - there is no issue with
implementing the TX path with scatterlist approach, but leaving
out the TX path makes it easier to review.
4. Functionality is tested with this proposal, but I have not conducted
perf testing yet. I'm not sure there are regressions, but I removed
perf claims from the cover letter until they can be re-confirmed.
5. Added Signed-off-by: contributors to the implementation.
6. Fixed some bugs with the RX path since RFC v1.
Any feedback welcome, but specifically the biggest pending questions
needing feedback IMO are:
1. Feedback on the scatterlist-based approach in general.
2. Netlink API (Patch 1 & 2).
3. Approach to handle all the drivers that expect to receive pages from
the page pool (Patch 6).
[1] https://lore.kernel.org/netdev/dfe4bae7-13a0-3c5d-d671-f61b375cb0b4@gmail.c…
[2] https://lore.kernel.org/netdev/CAHS8izPm6XRS54LdCDZVd0C75tA1zHSu6jLVO8nzTLX…
----------------------
* TL;DR:
Device memory TCP (devmem TCP) is a proposal for transferring data to and/or
from device memory efficiently, without bouncing the data to a host memory
buffer.
* Problem:
A large amount of data transfers have device memory as the source and/or
destination. Accelerators drastically increased the volume of such transfers.
Some examples include:
- ML accelerators transferring large amounts of training data from storage into
GPU/TPU memory. In some cases ML training setup time can be as long as 50% of
TPU compute time, improving data transfer throughput & efficiency can help
improving GPU/TPU utilization.
- Distributed training, where ML accelerators, such as GPUs on different hosts,
exchange data among them.
- Distributed raw block storage applications transfer large amounts of data with
remote SSDs, much of this data does not require host processing.
Today, the majority of the Device-to-Device data transfers the network are
implemented as the following low level operations: Device-to-Host copy,
Host-to-Host network transfer, and Host-to-Device copy.
The implementation is suboptimal, especially for bulk data transfers, and can
put significant strains on system resources, such as host memory bandwidth,
PCIe bandwidth, etc. One important reason behind the current state is the
kernel’s lack of semantics to express device to network transfers.
* Proposal:
In this patch series we attempt to optimize this use case by implementing
socket APIs that enable the user to:
1. send device memory across the network directly, and
2. receive incoming network packets directly into device memory.
Packet _payloads_ go directly from the NIC to device memory for receive and from
device memory to NIC for transmit.
Packet _headers_ go to/from host memory and are processed by the TCP/IP stack
normally. The NIC _must_ support header split to achieve this.
Advantages:
- Alleviate host memory bandwidth pressure, compared to existing
network-transfer + device-copy semantics.
- Alleviate PCIe BW pressure, by limiting data transfer to the lowest level
of the PCIe tree, compared to traditional path which sends data through the
root complex.
* Patch overview:
** Part 1: netlink API
Gives user ability to bind dma-buf to an RX queue.
** Part 2: scatterlist support
Currently the standard for device memory sharing is DMABUF, which doesn't
generate struct pages. On the other hand, networking stack (skbs, drivers, and
page pool) operate on pages. We have 2 options:
1. Generate struct pages for dmabuf device memory, or,
2. Modify the networking stack to process scatterlist.
Approach #1 was attempted in RFC v1. RFC v2 implements approach #2.
** part 3: page pool support
We piggy back on page pool memory providers proposal:
https://github.com/kuba-moo/linux/tree/pp-providers
It allows the page pool to define a memory provider that provides the
page allocation and freeing. It helps abstract most of the device memory
TCP changes from the driver.
** part 4: support for unreadable skb frags
Page pool iovs are not accessible by the host; we implement changes
throughput the networking stack to correctly handle skbs with unreadable
frags.
** Part 5: recvmsg() APIs
We define user APIs for the user to send and receive device memory.
Not included with this RFC is the GVE devmem TCP support, just to
simplify the review. Code available here if desired:
https://github.com/mina/linux/tree/tcpdevmem
This RFC is built on top of net-next with Jakub's pp-providers changes
cherry-picked.
* NIC dependencies:
1. (strict) Devmem TCP require the NIC to support header split, i.e. the
capability to split incoming packets into a header + payload and to put
each into a separate buffer. Devmem TCP works by using device memory
for the packet payload, and host memory for the packet headers.
2. (optional) Devmem TCP works better with flow steering support & RSS support,
i.e. the NIC's ability to steer flows into certain rx queues. This allows the
sysadmin to enable devmem TCP on a subset of the rx queues, and steer
devmem TCP traffic onto these queues and non devmem TCP elsewhere.
The NIC I have access to with these properties is the GVE with DQO support
running in Google Cloud, but any NIC that supports these features would suffice.
I may be able to help reviewers bring up devmem TCP on their NICs.
* Testing:
The series includes a udmabuf kselftest that show a simple use case of
devmem TCP and validates the entire data path end to end without
a dependency on a specific dmabuf provider.
** Test Setup
Kernel: net-next with this RFC and memory provider API cherry-picked
locally.
Hardware: Google Cloud A3 VMs.
NIC: GVE with header split & RSS & flow steering support.
Jakub Kicinski (2):
net: page_pool: factor out releasing DMA from releasing the page
net: page_pool: create hooks for custom page providers
Mina Almasry (14):
queue_api: define queue api
gve: implement queue api
net: netdev netlink api to bind dma-buf to a net device
netdev: support binding dma-buf to netdevice
netdev: netdevice devmem allocator
memory-provider: dmabuf devmem memory provider
page_pool: device memory support
page_pool: don't release iov on elevanted refcount
net: support non paged skb frags
net: add support for skbs with unreadable frags
tcp: RX path for devmem TCP
net: add SO_DEVMEM_DONTNEED setsockopt to release RX frags
net: add devmem TCP documentation
selftests: add ncdevmem, netcat for devmem TCP
Documentation/netlink/specs/netdev.yaml | 52 ++
Documentation/networking/devmem.rst | 270 ++++++++++
drivers/net/ethernet/google/gve/gve_adminq.c | 6 +-
drivers/net/ethernet/google/gve/gve_adminq.h | 3 +
drivers/net/ethernet/google/gve/gve_dqo.h | 2 +
drivers/net/ethernet/google/gve/gve_main.c | 286 +++++++++++
drivers/net/ethernet/google/gve/gve_rx_dqo.c | 5 +-
include/linux/netdevice.h | 24 +
include/linux/skbuff.h | 56 ++-
include/linux/socket.h | 1 +
include/net/devmem.h | 109 +++++
include/net/netdev_rx_queue.h | 1 +
include/net/page_pool/helpers.h | 162 +++++-
include/net/page_pool/types.h | 48 ++
include/net/sock.h | 2 +
include/net/tcp.h | 5 +-
include/uapi/asm-generic/socket.h | 6 +
include/uapi/linux/netdev.h | 19 +
include/uapi/linux/uio.h | 14 +
net/core/datagram.c | 6 +
net/core/dev.c | 314 +++++++++++-
net/core/gro.c | 7 +-
net/core/netdev-genl-gen.c | 19 +
net/core/netdev-genl-gen.h | 2 +
net/core/netdev-genl.c | 124 +++++
net/core/page_pool.c | 239 +++++++--
net/core/skbuff.c | 108 +++-
net/core/sock.c | 38 ++
net/ipv4/tcp.c | 196 +++++++-
net/ipv4/tcp_input.c | 13 +-
net/ipv4/tcp_ipv4.c | 8 +
net/ipv4/tcp_output.c | 5 +-
net/packet/af_packet.c | 4 +-
tools/include/uapi/linux/netdev.h | 19 +
tools/testing/selftests/net/.gitignore | 1 +
tools/testing/selftests/net/Makefile | 5 +
tools/testing/selftests/net/ncdevmem.c | 489 +++++++++++++++++++
37 files changed, 2585 insertions(+), 83 deletions(-)
create mode 100644 Documentation/networking/devmem.rst
create mode 100644 include/net/devmem.h
create mode 100644 tools/testing/selftests/net/ncdevmem.c
--
2.43.0.472.g3155946c3a-goog
Major changes in RFC v5:
------------------------
1. Rebased on top of 'Abstract page from net stack' series and used the
new netmem type to refer to LSB set pointers instead of re-using
struct page.
2. Downgraded this series back to RFC and called it RFC v5. This is
because this series is now dependent on 'Abstract page from net
stack'[1] and the queue API. Both are removed from the series to
pre-requisite work.
3. Reworked the page_pool devmem support to use netmem and for some
more unified handling.
4. Reworked the reference counting of net_iov (renamed from
page_pool_iov) to use pp_ref_count for refcounting.
The full changes including the dependent series and GVE page pool
support is here:
https://github.com/mina/linux/commits/tcpdevmem-rfcv5/
[1] https://patchwork.kernel.org/project/netdevbpf/list/?series=810774
Major changes in v1:
--------------------
1. Implemented MVP queue API ndos to remove the userspace-visible
driver reset.
2. Fixed issues in the napi_pp_put_page() devmem frag unref path.
3. Removed RFC tag.
Many smaller addressed comments across all the patches (patches have
individual change log).
Full tree including the rest of the GVE driver changes:
https://github.com/mina/linux/commits/tcpdevmem-v1
Changes in RFC v3:
------------------
1. Pulled in the memory-provider dependency from Jakub's RFC[1] to make the
series reviewable and mergable.
2. Implemented multi-rx-queue binding which was a todo in v2.
3. Fix to cmsg handling.
The sticking point in RFC v2[2] was the device reset required to refill
the device rx-queues after the dmabuf bind/unbind. The solution
suggested as I understand is a subset of the per-queue management ops
Jakub suggested or similar:
https://lore.kernel.org/netdev/20230815171638.4c057dcd@kernel.org/
This is not addressed in this revision, because:
1. This point was discussed at netconf & netdev and there is openness to
using the current approach of requiring a device reset.
2. Implementing individual queue resetting seems to be difficult for my
test bed with GVE. My prototype to test this ran into issues with the
rx-queues not coming back up properly if reset individually. At the
moment I'm unsure if it's a mistake in the POC or a genuine issue in
the virtualization stack behind GVE, which currently doesn't test
individual rx-queue restart.
3. Our usecases are not bothered by requiring a device reset to refill
the buffer queues, and we'd like to support NICs that run into this
limitation with resetting individual queues.
My thought is that drivers that have trouble with per-queue configs can
use the support in this series, while drivers that support new netdev
ops to reset individual queues can automatically reset the queue as
part of the dma-buf bind/unbind.
The same approach with device resets is presented again for consideration
with other sticking points addressed.
This proposal includes the rx devmem path only proposed for merge. For a
snapshot of my entire tree which includes the GVE POC page pool support &
device memory support:
https://github.com/torvalds/linux/compare/master...mina:linux:tcpdevmem-v3
[1] https://lore.kernel.org/netdev/f8270765-a27b-6ccf-33ea-cda097168d79@redhat.…
[2] https://lore.kernel.org/netdev/CAHS8izOVJGJH5WF68OsRWFKJid1_huzzUK+hpKbLcL4…
Changes in RFC v2:
------------------
The sticking point in RFC v1[1] was the dma-buf pages approach we used to
deliver the device memory to the TCP stack. RFC v2 is a proof-of-concept
that attempts to resolve this by implementing scatterlist support in the
networking stack, such that we can import the dma-buf scatterlist
directly. This is the approach proposed at a high level here[2].
Detailed changes:
1. Replaced dma-buf pages approach with importing scatterlist into the
page pool.
2. Replace the dma-buf pages centric API with a netlink API.
3. Removed the TX path implementation - there is no issue with
implementing the TX path with scatterlist approach, but leaving
out the TX path makes it easier to review.
4. Functionality is tested with this proposal, but I have not conducted
perf testing yet. I'm not sure there are regressions, but I removed
perf claims from the cover letter until they can be re-confirmed.
5. Added Signed-off-by: contributors to the implementation.
6. Fixed some bugs with the RX path since RFC v1.
Any feedback welcome, but specifically the biggest pending questions
needing feedback IMO are:
1. Feedback on the scatterlist-based approach in general.
2. Netlink API (Patch 1 & 2).
3. Approach to handle all the drivers that expect to receive pages from
the page pool (Patch 6).
[1] https://lore.kernel.org/netdev/dfe4bae7-13a0-3c5d-d671-f61b375cb0b4@gmail.c…
[2] https://lore.kernel.org/netdev/CAHS8izPm6XRS54LdCDZVd0C75tA1zHSu6jLVO8nzTLX…
----------------------
* TL;DR:
Device memory TCP (devmem TCP) is a proposal for transferring data to and/or
from device memory efficiently, without bouncing the data to a host memory
buffer.
* Problem:
A large amount of data transfers have device memory as the source and/or
destination. Accelerators drastically increased the volume of such transfers.
Some examples include:
- ML accelerators transferring large amounts of training data from storage into
GPU/TPU memory. In some cases ML training setup time can be as long as 50% of
TPU compute time, improving data transfer throughput & efficiency can help
improving GPU/TPU utilization.
- Distributed training, where ML accelerators, such as GPUs on different hosts,
exchange data among them.
- Distributed raw block storage applications transfer large amounts of data with
remote SSDs, much of this data does not require host processing.
Today, the majority of the Device-to-Device data transfers the network are
implemented as the following low level operations: Device-to-Host copy,
Host-to-Host network transfer, and Host-to-Device copy.
The implementation is suboptimal, especially for bulk data transfers, and can
put significant strains on system resources, such as host memory bandwidth,
PCIe bandwidth, etc. One important reason behind the current state is the
kernel’s lack of semantics to express device to network transfers.
* Proposal:
In this patch series we attempt to optimize this use case by implementing
socket APIs that enable the user to:
1. send device memory across the network directly, and
2. receive incoming network packets directly into device memory.
Packet _payloads_ go directly from the NIC to device memory for receive and from
device memory to NIC for transmit.
Packet _headers_ go to/from host memory and are processed by the TCP/IP stack
normally. The NIC _must_ support header split to achieve this.
Advantages:
- Alleviate host memory bandwidth pressure, compared to existing
network-transfer + device-copy semantics.
- Alleviate PCIe BW pressure, by limiting data transfer to the lowest level
of the PCIe tree, compared to traditional path which sends data through the
root complex.
* Patch overview:
** Part 1: netlink API
Gives user ability to bind dma-buf to an RX queue.
** Part 2: scatterlist support
Currently the standard for device memory sharing is DMABUF, which doesn't
generate struct pages. On the other hand, networking stack (skbs, drivers, and
page pool) operate on pages. We have 2 options:
1. Generate struct pages for dmabuf device memory, or,
2. Modify the networking stack to process scatterlist.
** part 3: page pool support
We piggy back on page pool memory providers proposal:
https://github.com/kuba-moo/linux/tree/pp-providers
It allows the page pool to define a memory provider that provides the
page allocation and freeing. It helps abstract most of the device memory
TCP changes from the driver.
** part 4: support for unreadable skb frags
Page pool iovs are not accessible by the host; we implement changes
throughput the networking stack to correctly handle skbs with unreadable
frags.
** Part 5: recvmsg() APIs
We define user APIs for the user to send and receive device memory.
Not included with this series is the GVE devmem TCP support, just to
simplify the review. Code available here if desired:
https://github.com/mina/linux/tree/tcpdevmem
This series is built on top of net-next with Jakub's pp-providers changes
cherry-picked.
* NIC dependencies:
1. (strict) Devmem TCP require the NIC to support header split, i.e. the
capability to split incoming packets into a header + payload and to put
each into a separate buffer. Devmem TCP works by using device memory
for the packet payload, and host memory for the packet headers.
2. (optional) Devmem TCP works better with flow steering support & RSS support,
i.e. the NIC's ability to steer flows into certain rx queues. This allows the
sysadmin to enable devmem TCP on a subset of the rx queues, and steer
devmem TCP traffic onto these queues and non devmem TCP elsewhere.
The NIC I have access to with these properties is the GVE with DQO support
running in Google Cloud, but any NIC that supports these features would suffice.
I may be able to help reviewers bring up devmem TCP on their NICs.
* Testing:
The series includes a udmabuf kselftest that show a simple use case of
devmem TCP and validates the entire data path end to end without
a dependency on a specific dmabuf provider.
** Test Setup
Kernel: net-next with this series and memory provider API cherry-picked
locally.
Hardware: Google Cloud A3 VMs.
NIC: GVE with header split & RSS & flow steering support.
Cc: Pavel Begunkov <asml.silence(a)gmail.com>
Cc: David Wei <dw(a)davidwei.uk>
Cc: Jason Gunthorpe <jgg(a)ziepe.ca>
Cc: Yunsheng Lin <linyunsheng(a)huawei.com>
Cc: Shailend Chand <shailend(a)google.com>
Cc: Harshitha Ramamurthy <hramamurthy(a)google.com>
Cc: Shakeel Butt <shakeelb(a)google.com>
Cc: Jeroen de Borst <jeroendb(a)google.com>
Cc: Praveen Kaligineedi <pkaligineedi(a)google.com>
Jakub Kicinski (1):
net: page_pool: create hooks for custom page providers
Mina Almasry (13):
net: page_pool: factor out page_pool recycle check
net: netdev netlink api to bind dma-buf to a net device
netdev: support binding dma-buf to netdevice
netdev: netdevice devmem allocator
page_pool: convert to use netmem
page_pool: devmem support
memory-provider: dmabuf devmem memory provider
net: support non paged skb frags
net: add support for skbs with unreadable frags
tcp: RX path for devmem TCP
net: add SO_DEVMEM_DONTNEED setsockopt to release RX frags
net: add devmem TCP documentation
selftests: add ncdevmem, netcat for devmem TCP
Documentation/netlink/specs/netdev.yaml | 52 +++
Documentation/networking/devmem.rst | 271 +++++++++++++
Documentation/networking/index.rst | 1 +
arch/alpha/include/uapi/asm/socket.h | 8 +-
arch/mips/include/uapi/asm/socket.h | 6 +
arch/parisc/include/uapi/asm/socket.h | 6 +
arch/sparc/include/uapi/asm/socket.h | 6 +
include/linux/skbuff.h | 68 +++-
include/linux/socket.h | 1 +
include/net/devmem.h | 127 ++++++
include/net/netdev_rx_queue.h | 1 +
include/net/netmem.h | 223 ++++++++++-
include/net/page_pool/helpers.h | 117 ++++--
include/net/page_pool/types.h | 27 +-
include/net/sock.h | 2 +
include/net/tcp.h | 5 +-
include/trace/events/page_pool.h | 28 +-
include/uapi/asm-generic/socket.h | 6 +
include/uapi/linux/netdev.h | 19 +
include/uapi/linux/uio.h | 14 +
net/bpf/test_run.c | 5 +-
net/core/datagram.c | 6 +
net/core/dev.c | 314 ++++++++++++++-
net/core/gro.c | 7 +-
net/core/netdev-genl-gen.c | 19 +
net/core/netdev-genl-gen.h | 2 +
net/core/netdev-genl.c | 123 ++++++
net/core/page_pool.c | 419 +++++++++++++-------
net/core/skbuff.c | 92 ++++-
net/core/sock.c | 45 +++
net/ipv4/tcp.c | 196 +++++++++-
net/ipv4/tcp_input.c | 13 +-
net/ipv4/tcp_ipv4.c | 9 +
net/ipv4/tcp_output.c | 5 +-
net/packet/af_packet.c | 4 +-
tools/include/uapi/linux/netdev.h | 19 +
tools/testing/selftests/net/.gitignore | 1 +
tools/testing/selftests/net/Makefile | 5 +
tools/testing/selftests/net/ncdevmem.c | 489 ++++++++++++++++++++++++
39 files changed, 2527 insertions(+), 234 deletions(-)
create mode 100644 Documentation/networking/devmem.rst
create mode 100644 include/net/devmem.h
create mode 100644 tools/testing/selftests/net/ncdevmem.c
--
2.43.0.472.g3155946c3a-goog
From: Zi Yan <ziy(a)nvidia.com>
Hi all,
File folio supports any order and people would like to support flexible orders
for anonymous folio[1] too. Currently, split_huge_page() only splits a huge
page to order-0 pages, but splitting to orders higher than 0 is also useful.
This patchset adds support for splitting a huge page to any lower order pages
and uses it during file folio truncate operations.
The patchset is on top of mm-everything-2023-03-27-21-20.
Changelog
===
Since v2
---
1. Fixed an issue in __split_page_owner() introduced during my rebase
Since v1
---
1. Changed split_page_memcg() and split_page_owner() parameter to use order
2. Used folio_test_pmd_mappable() in place of the equivalent code
Details
===
* Patch 1 changes split_page_memcg() to use order instead of nr_pages
* Patch 2 changes split_page_owner() to use order instead of nr_pages
* Patch 3 and 4 add new_order parameter split_page_memcg() and
split_page_owner() and prepare for upcoming changes.
* Patch 5 adds split_huge_page_to_list_to_order() to split a huge page
to any lower order. The original split_huge_page_to_list() calls
split_huge_page_to_list_to_order() with new_order = 0.
* Patch 6 uses split_huge_page_to_list_to_order() in large pagecache folio
truncation instead of split the large folio all the way down to order-0.
* Patch 7 adds a test API to debugfs and test cases in
split_huge_page_test selftests.
Comments and/or suggestions are welcome.
[1] https://lore.kernel.org/linux-mm/Y%2FblF0GIunm+pRIC@casper.infradead.org/
Zi Yan (7):
mm/memcg: use order instead of nr in split_page_memcg()
mm/page_owner: use order instead of nr in split_page_owner()
mm: memcg: make memcg huge page split support any order split.
mm: page_owner: add support for splitting to any order in split
page_owner.
mm: thp: split huge page to any lower order pages.
mm: truncate: split huge page cache page to a non-zero order if
possible.
mm: huge_memory: enable debugfs to split huge pages to any order.
include/linux/huge_mm.h | 10 +-
include/linux/memcontrol.h | 4 +-
include/linux/page_owner.h | 10 +-
mm/huge_memory.c | 137 ++++++++---
mm/memcontrol.c | 10 +-
mm/page_alloc.c | 8 +-
mm/page_owner.c | 8 +-
mm/truncate.c | 21 +-
.../selftests/mm/split_huge_page_test.c | 225 +++++++++++++++++-
9 files changed, 365 insertions(+), 68 deletions(-)
--
2.39.2
Here's a follow-up from my RFC series last year:
https://lore.kernel.org/lkml/20221004093131.40392-1-thuth@redhat.com/T/
and from v1 earlier this year:
https://lore.kernel.org/kvm/20230712075910.22480-1-thuth@redhat.com/
Basic idea of this series is now to use the kselftest_harness.h
framework to get TAP output in the tests, so that it is easier
for the user to see what is going on, and e.g. to be able to
detect whether a certain test is part of the test binary or not
(which is useful when tests get extended in the course of time).
v2:
- Dropped the "Rename the ASSERT_EQ macro" patch (already merged)
- Split the fixes in the sync_regs_test into separate patches
(see the first two patches)
- Introduce the KVM_ONE_VCPU_TEST_SUITE() macro as suggested
by Sean (see third patch) and use it in the following patches
- Add a new patch to convert vmx_pmu_caps_test.c, too
Thomas Huth (7):
KVM: selftests: x86: sync_regs_test: Use vcpu_run() where appropriate
KVM: selftests: x86: sync_regs_test: Get regs structure before
modifying it
KVM: selftests: Add a macro to define a test with one vcpu
KVM: selftests: x86: Use TAP interface in the sync_regs test
KVM: selftests: x86: Use TAP interface in the fix_hypercall test
KVM: selftests: x86: Use TAP interface in the vmx_pmu_caps test
KVM: selftests: x86: Use TAP interface in the userspace_msr_exit test
.../selftests/kvm/include/kvm_test_harness.h | 35 +++++
.../selftests/kvm/x86_64/fix_hypercall_test.c | 27 ++--
.../selftests/kvm/x86_64/sync_regs_test.c | 121 +++++++++++++-----
.../kvm/x86_64/userspace_msr_exit_test.c | 19 +--
.../selftests/kvm/x86_64/vmx_pmu_caps_test.c | 50 ++------
5 files changed, 160 insertions(+), 92 deletions(-)
create mode 100644 tools/testing/selftests/kvm/include/kvm_test_harness.h
--
2.41.0
Other mechanisms for querying the peak memory usage of either a process
or v1 memory cgroup allow for resetting the high watermark. Restore
parity with those mechanisms.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets
the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc
directory resets the peak RSS.
This change copies the cgroup v1 behavior so any write to the
memory.peak and memory.swap.peak pseudo-files reset the high watermark
to the current usage.
This behavior is particularly useful for work scheduling systems that
need to track memory usage of worker processes/cgroups per-work-item.
Since memory can't be squeezed like CPU can (the OOM-killer has
opinions), these systems need to track the peak memory usage to compute
system/container fullness when binpacking workitems.
Signed-off-by: David Finkel <davidf(a)vimeo.com>
---
Documentation/admin-guide/cgroup-v2.rst | 20 +++---
mm/memcontrol.c | 23 ++++++
.../selftests/cgroup/test_memcontrol.c | 72 ++++++++++++++++---
3 files changed, 99 insertions(+), 16 deletions(-)
diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst
index 3f85254f3cef..95af0628dc44 100644
--- a/Documentation/admin-guide/cgroup-v2.rst
+++ b/Documentation/admin-guide/cgroup-v2.rst
@@ -1305,11 +1305,13 @@ PAGE_SIZE multiple when read back.
reclaim induced by memory.reclaim.
memory.peak
- A read-only single value file which exists on non-root
- cgroups.
+ A read-write single value file which exists on non-root cgroups.
+
+ The max memory usage recorded for the cgroup and its descendants since
+ either the creation of the cgroup or the most recent reset.
- The max memory usage recorded for the cgroup and its
- descendants since the creation of the cgroup.
+ Any non-empty write to this file resets it to the current memory usage.
+ All content written is completely ignored.
memory.oom.group
A read-write single value file which exists on non-root
@@ -1626,11 +1628,13 @@ PAGE_SIZE multiple when read back.
Healthy workloads are not expected to reach this limit.
memory.swap.peak
- A read-only single value file which exists on non-root
- cgroups.
+ A read-write single value file which exists on non-root cgroups.
+
+ The max swap usage recorded for the cgroup and its descendants since
+ the creation of the cgroup or the most recent reset.
- The max swap usage recorded for the cgroup and its
- descendants since the creation of the cgroup.
+ Any non-empty write to this file resets it to the current swap usage.
+ All content written is completely ignored.
memory.swap.max
A read-write single value file which exists on non-root
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 1c1061df9cd1..b04af158922d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -25,6 +25,7 @@
* Copyright (C) 2020 Alibaba, Inc, Alex Shi
*/
+#include <linux/cgroup-defs.h>
#include <linux/page_counter.h>
#include <linux/memcontrol.h>
#include <linux/cgroup.h>
@@ -6635,6 +6636,16 @@ static u64 memory_peak_read(struct cgroup_subsys_state *css,
return (u64)memcg->memory.watermark * PAGE_SIZE;
}
+static ssize_t memory_peak_write(struct kernfs_open_file *of,
+ char *buf, size_t nbytes, loff_t off)
+{
+ struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
+
+ page_counter_reset_watermark(&memcg->memory);
+
+ return nbytes;
+}
+
static int memory_min_show(struct seq_file *m, void *v)
{
return seq_puts_memcg_tunable(m,
@@ -6947,6 +6958,7 @@ static struct cftype memory_files[] = {
.name = "peak",
.flags = CFTYPE_NOT_ON_ROOT,
.read_u64 = memory_peak_read,
+ .write = memory_peak_write,
},
{
.name = "min",
@@ -7917,6 +7929,16 @@ static u64 swap_peak_read(struct cgroup_subsys_state *css,
return (u64)memcg->swap.watermark * PAGE_SIZE;
}
+static ssize_t swap_peak_write(struct kernfs_open_file *of,
+ char *buf, size_t nbytes, loff_t off)
+{
+ struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
+
+ page_counter_reset_watermark(&memcg->swap);
+
+ return nbytes;
+}
+
static int swap_high_show(struct seq_file *m, void *v)
{
return seq_puts_memcg_tunable(m,
@@ -7999,6 +8021,7 @@ static struct cftype swap_files[] = {
.name = "swap.peak",
.flags = CFTYPE_NOT_ON_ROOT,
.read_u64 = swap_peak_read,
+ .write = swap_peak_write,
},
{
.name = "swap.events",
diff --git a/tools/testing/selftests/cgroup/test_memcontrol.c b/tools/testing/selftests/cgroup/test_memcontrol.c
index c7c9572003a8..0326c317f1f2 100644
--- a/tools/testing/selftests/cgroup/test_memcontrol.c
+++ b/tools/testing/selftests/cgroup/test_memcontrol.c
@@ -161,12 +161,12 @@ static int alloc_pagecache_50M_check(const char *cgroup, void *arg)
/*
* This test create a memory cgroup, allocates
* some anonymous memory and some pagecache
- * and check memory.current and some memory.stat values.
+ * and checks memory.current, memory.peak, and some memory.stat values.
*/
-static int test_memcg_current(const char *root)
+static int test_memcg_current_peak(const char *root)
{
int ret = KSFT_FAIL;
- long current;
+ long current, peak, peak_reset;
char *memcg;
memcg = cg_name(root, "memcg_test");
@@ -180,12 +180,32 @@ static int test_memcg_current(const char *root)
if (current != 0)
goto cleanup;
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak != 0)
+ goto cleanup;
+
if (cg_run(memcg, alloc_anon_50M_check, NULL))
goto cleanup;
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak < MB(50))
+ goto cleanup;
+
+ peak_reset = cg_write(memcg, "memory.peak", "\n");
+ if (peak_reset != 0)
+ goto cleanup;
+
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak > MB(30))
+ goto cleanup;
+
if (cg_run(memcg, alloc_pagecache_50M_check, NULL))
goto cleanup;
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak < MB(50))
+ goto cleanup;
+
ret = KSFT_PASS;
cleanup:
@@ -815,13 +835,14 @@ static int alloc_anon_50M_check_swap(const char *cgroup, void *arg)
/*
* This test checks that memory.swap.max limits the amount of
- * anonymous memory which can be swapped out.
+ * anonymous memory which can be swapped out. Additionally, it verifies that
+ * memory.swap.peak reflects the high watermark and can be reset.
*/
-static int test_memcg_swap_max(const char *root)
+static int test_memcg_swap_max_peak(const char *root)
{
int ret = KSFT_FAIL;
char *memcg;
- long max;
+ long max, peak;
if (!is_swap_enabled())
return KSFT_SKIP;
@@ -838,6 +859,12 @@ static int test_memcg_swap_max(const char *root)
goto cleanup;
}
+ if (cg_read_long(memcg, "memory.swap.peak"))
+ goto cleanup;
+
+ if (cg_read_long(memcg, "memory.peak"))
+ goto cleanup;
+
if (cg_read_strcmp(memcg, "memory.max", "max\n"))
goto cleanup;
@@ -860,6 +887,27 @@ static int test_memcg_swap_max(const char *root)
if (cg_read_key_long(memcg, "memory.events", "oom_kill ") != 1)
goto cleanup;
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak < MB(29))
+ goto cleanup;
+
+ peak = cg_read_long(memcg, "memory.swap.peak");
+ if (peak < MB(29))
+ goto cleanup;
+
+ if (cg_write(memcg, "memory.swap.peak", "\n"))
+ goto cleanup;
+
+ if (cg_read_long(memcg, "memory.swap.peak") > MB(10))
+ goto cleanup;
+
+
+ if (cg_write(memcg, "memory.peak", "\n"))
+ goto cleanup;
+
+ if (cg_read_long(memcg, "memory.peak"))
+ goto cleanup;
+
if (cg_run(memcg, alloc_anon_50M_check_swap, (void *)MB(30)))
goto cleanup;
@@ -867,6 +915,14 @@ static int test_memcg_swap_max(const char *root)
if (max <= 0)
goto cleanup;
+ peak = cg_read_long(memcg, "memory.peak");
+ if (peak < MB(29))
+ goto cleanup;
+
+ peak = cg_read_long(memcg, "memory.swap.peak");
+ if (peak < MB(19))
+ goto cleanup;
+
ret = KSFT_PASS;
cleanup:
@@ -1293,7 +1349,7 @@ struct memcg_test {
const char *name;
} tests[] = {
T(test_memcg_subtree_control),
- T(test_memcg_current),
+ T(test_memcg_current_peak),
T(test_memcg_min),
T(test_memcg_low),
T(test_memcg_high),
@@ -1301,7 +1357,7 @@ struct memcg_test {
T(test_memcg_max),
T(test_memcg_reclaim),
T(test_memcg_oom_events),
- T(test_memcg_swap_max),
+ T(test_memcg_swap_max_peak),
T(test_memcg_sock),
T(test_memcg_oom_group_leaf_events),
T(test_memcg_oom_group_parent_events),
--
2.39.2
On Mon, Nov 27, 2023 at 11:49:16AM +0000, Felix Huettner wrote:
> conntrack zones are heavily used by tools like openvswitch to run
> multiple virtual "routers" on a single machine. In this context each
> conntrack zone matches to a single router, thereby preventing
> overlapping IPs from becoming issues.
> In these systems it is common to operate on all conntrack entries of a
> given zone, e.g. to delete them when a router is deleted. Previously this
> required these tools to dump the full conntrack table and filter out the
> relevant entries in userspace potentially causing performance issues.
>
> To do this we reuse the existing CTA_ZONE attribute. This was previous
> parsed but not used during dump and flush requests. Now if CTA_ZONE is
> set we filter these operations based on the provided zone.
> However this means that users that previously passed CTA_ZONE will
> experience a difference in functionality.
>
> Alternatively CTA_FILTER could have been used for the same
> functionality. However it is not yet supported during flush requests and
> is only available when using AF_INET or AF_INET6.
For the record, this is applied to nf-next.
The kernel has recently added support for shadow stacks, currently
x86 only using their CET feature but both arm64 and RISC-V have
equivalent features (GCS and Zicfiss respectively), I am actively
working on GCS[1]. With shadow stacks the hardware maintains an
additional stack containing only the return addresses for branch
instructions which is not generally writeable by userspace and ensures
that any returns are to the recorded addresses. This provides some
protection against ROP attacks and making it easier to collect call
stacks. These shadow stacks are allocated in the address space of the
userspace process.
Our API for shadow stacks does not currently offer userspace any
flexiblity for managing the allocation of shadow stacks for newly
created threads, instead the kernel allocates a new shadow stack with
the same size as the normal stack whenever a thread is created with the
feature enabled. The stacks allocated in this way are freed by the
kernel when the thread exits or shadow stacks are disabled for the
thread. This lack of flexibility and control isn't ideal, in the vast
majority of cases the shadow stack will be over allocated and the
implicit allocation and deallocation is not consistent with other
interfaces. As far as I can tell the interface is done in this manner
mainly because the shadow stack patches were in development since before
clone3() was implemented.
Since clone3() is readily extensible let's add support for specifying a
shadow stack when creating a new thread or process in a similar manner
to how the normal stack is specified, keeping the current implicit
allocation behaviour if one is not specified either with clone3() or
through the use of clone(). Unlike normal stacks only the shadow stack
size is specified, similar issues to those that lead to the creation of
map_shadow_stack() apply.
Please note that the x86 portions of this code are build tested only, I
don't appear to have a system that can run CET avaible to me, I have
done testing with an integration into my pending work for GCS. There is
some possibility that the arm64 implementation may require the use of
clone3() and explicit userspace allocation of shadow stacks, this is
still under discussion.
A new architecture feature Kconfig option for shadow stacks is added as
here, this was suggested as part of the review comments for the arm64
GCS series and since we need to detect if shadow stacks are supported it
seemed sensible to roll it in here.
[1] https://lore.kernel.org/r/20231009-arm64-gcs-v6-0-78e55deaa4dd@kernel.org/
Signed-off-by: Mark Brown <broonie(a)kernel.org>
---
Changes in v3:
- Rebase onto v6.7-rc2.
- Remove stale shadow_stack in internal kargs.
- If a shadow stack is specified unconditionally use it regardless of
CLONE_ parameters.
- Force enable shadow stacks in the selftest.
- Update changelogs for RISC-V feature rename.
- Link to v2: https://lore.kernel.org/r/20231114-clone3-shadow-stack-v2-0-b613f8681155@ke…
Changes in v2:
- Rebase onto v6.7-rc1.
- Remove ability to provide preallocated shadow stack, just specify the
desired size.
- Link to v1: https://lore.kernel.org/r/20231023-clone3-shadow-stack-v1-0-d867d0b5d4d0@ke…
---
Mark Brown (5):
mm: Introduce ARCH_HAS_USER_SHADOW_STACK
fork: Add shadow stack support to clone3()
selftests/clone3: Factor more of main loop into test_clone3()
selftests/clone3: Allow tests to flag if -E2BIG is a valid error code
kselftest/clone3: Test shadow stack support
arch/x86/Kconfig | 1 +
arch/x86/include/asm/shstk.h | 11 +-
arch/x86/kernel/process.c | 2 +-
arch/x86/kernel/shstk.c | 59 +++++--
fs/proc/task_mmu.c | 2 +-
include/linux/mm.h | 2 +-
include/linux/sched/task.h | 1 +
include/uapi/linux/sched.h | 4 +
kernel/fork.c | 22 ++-
mm/Kconfig | 6 +
tools/testing/selftests/clone3/clone3.c | 200 +++++++++++++++++-----
tools/testing/selftests/clone3/clone3_selftests.h | 7 +
12 files changed, 250 insertions(+), 67 deletions(-)
---
base-commit: 98b1cc82c4affc16f5598d4fa14b1858671b2263
change-id: 20231019-clone3-shadow-stack-15d40d2bf536
Best regards,
--
Mark Brown <broonie(a)kernel.org>
When execute the dirty_log_test on some aarch64 machine, it sometimes
trigger the ASSERT:
==== Test Assertion Failure ====
dirty_log_test.c:384: dirty_ring_vcpu_ring_full
pid=14854 tid=14854 errno=22 - Invalid argument
1 0x00000000004033eb: dirty_ring_collect_dirty_pages at dirty_log_test.c:384
2 0x0000000000402d27: log_mode_collect_dirty_pages at dirty_log_test.c:505
3 (inlined by) run_test at dirty_log_test.c:802
4 0x0000000000403dc7: for_each_guest_mode at guest_modes.c:100
5 0x0000000000401dff: main at dirty_log_test.c:941 (discriminator 3)
6 0x0000ffff9be173c7: ?? ??:0
7 0x0000ffff9be1749f: ?? ??:0
8 0x000000000040206f: _start at ??:?
Didn't continue vcpu even without ring full
The dirty_log_test fails when execute the dirty-ring test, this is
because the sem_vcpu_cont and the sem_vcpu_stop is non-zero value when
execute the dirty_ring_collect_dirty_pages() function. When those two
sem_t variables are non-zero, the dirty_ring_wait_vcpu() at the
beginning of the dirty_ring_collect_dirty_pages() will not wait for the
vcpu to stop, but continue to execute the following code. In this case,
before vcpu stop, if the dirty_ring_vcpu_ring_full is true, and the
dirty_ring_collect_dirty_pages() has passed the check for the
dirty_ring_vcpu_ring_full but hasn't execute the check for the
continued_vcpu, the vcpu stop, and set the dirty_ring_vcpu_ring_full to
false. Then dirty_ring_collect_dirty_pages() will trigger the ASSERT.
Why sem_vcpu_cont and sem_vcpu_stop can be non-zero value? It's because
the dirty_ring_before_vcpu_join() execute the sem_post(&sem_vcpu_cont)
at the end of each dirty-ring test. It can cause two cases:
1. sem_vcpu_cont be non-zero. When we set the host_quit to be true,
the vcpu_worker directly see the host_quit to be true, it quit. So
the log_mode_before_vcpu_join() function will set the sem_vcpu_cont
to 1, since the vcpu_worker has quit, it won't consume it.
2. sem_vcpu_stop be non-zero. When we set the host_quit to be true,
the vcpu_worker has entered the guest state, the next time it exit
from guest state, it will set the sem_vcpu_stop to 1, and then see
the host_quit, no one will consume the sem_vcpu_stop.
When execute more and more dirty-ring tests, the sem_vcpu_cont and
sem_vcpu_stop can be larger and larger, which makes many code paths
don't wait for the sem_t. Thus finally cause the problem.
To fix this problem, we can wait a while before set the host_quit to
true, which gives the vcpu time to enter the guest state, so it will
exit again. Then we can wait the vcpu to exit, and let it continue
again, then the vcpu will see the host_quit. Thus the sem_vcpu_cont and
sem_vcpu_stop will be both zero when test finished.
Signed-off-by: Shaoqin Huang <shahuang(a)redhat.com>
---
v1->v2:
- Fix the real logic bug, not just fresh the context.
v1: https://lore.kernel.org/all/20231116093536.22256-1-shahuang@redhat.com/
---
tools/testing/selftests/kvm/dirty_log_test.c | 16 +++++++++++++++-
1 file changed, 15 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/kvm/dirty_log_test.c b/tools/testing/selftests/kvm/dirty_log_test.c
index 936f3a8d1b83..a6e0ff46a07c 100644
--- a/tools/testing/selftests/kvm/dirty_log_test.c
+++ b/tools/testing/selftests/kvm/dirty_log_test.c
@@ -417,7 +417,8 @@ static void dirty_ring_after_vcpu_run(struct kvm_vcpu *vcpu, int ret, int err)
static void dirty_ring_before_vcpu_join(void)
{
- /* Kick another round of vcpu just to make sure it will quit */
+ /* Wait vcpu exit, and let it continue to see the host_quit. */
+ dirty_ring_wait_vcpu();
sem_post(&sem_vcpu_cont);
}
@@ -719,6 +720,7 @@ static void run_test(enum vm_guest_mode mode, void *arg)
struct kvm_vm *vm;
unsigned long *bmap;
uint32_t ring_buf_idx = 0;
+ int sem_val;
if (!log_mode_supported()) {
print_skip("Log mode '%s' not supported",
@@ -726,6 +728,11 @@ static void run_test(enum vm_guest_mode mode, void *arg)
return;
}
+ sem_getvalue(&sem_vcpu_stop, &sem_val);
+ assert(sem_val == 0);
+ sem_getvalue(&sem_vcpu_cont, &sem_val);
+ assert(sem_val == 0);
+
/*
* We reserve page table for 2 times of extra dirty mem which
* will definitely cover the original (1G+) test range. Here
@@ -825,6 +832,13 @@ static void run_test(enum vm_guest_mode mode, void *arg)
sync_global_to_guest(vm, iteration);
}
+ /*
+ *
+ * Before we set the host_quit, let the vcpu has time to run, to make
+ * sure we consume the sem_vcpu_stop and the vcpu consume the
+ * sem_vcpu_cont, to keep the semaphore balance.
+ */
+ usleep(p->interval * 1000);
/* Tell the vcpu thread to quit */
host_quit = true;
log_mode_before_vcpu_join();
--
2.40.1
This patch series give a proposal to support guest VM running
in user mode and in canonical linear address organization as
well.
First design to parition the 64-bit canonical linear address space
into two half parts belonging to user-mode and supervisor-mode
respectively, similar as the organization of linear addresses used
in linux OS. Currently the linear addresses use 48-bit canonical
format in which bits 63:47 of the address are identical.
Secondly setup page table mapping the same guest physical address
of test code and data segment onto both user-mode and supervisor-mode
address space. It allows guest in different runtime mode, i.e.
user or supervisor, can run one code base in the corresponding
linear address space.
Also provide the runtime environment setup API for switching to
user mode execution.
Zeng Guang (8):
KVM: selftests: x86: Fix bug in addr_arch_gva2gpa()
KVM: selftests: x86: Support guest running on canonical linear-address
organization
KVM: selftests: Add virt_arch_ucall_prealloc() arch specific
implementation
KVM : selftests : Adapt selftest cases to kernel canonical linear
address
KVM: selftests: x86: Prepare setup for user mode support
KVM: selftests: x86: Allow user to access user-mode address and I/O
address space
KVM: selftests: x86: Support vcpu run in user mode
KVM: selftests: x86: Add KVM forced emulation prefix capability
.../selftests/kvm/include/kvm_util_base.h | 20 ++-
.../selftests/kvm/include/x86_64/processor.h | 48 ++++++-
.../selftests/kvm/lib/aarch64/processor.c | 5 +
tools/testing/selftests/kvm/lib/kvm_util.c | 6 +-
.../selftests/kvm/lib/riscv/processor.c | 5 +
.../selftests/kvm/lib/s390x/processor.c | 5 +
.../testing/selftests/kvm/lib/ucall_common.c | 2 +
.../selftests/kvm/lib/x86_64/processor.c | 117 ++++++++++++++----
.../selftests/kvm/set_memory_region_test.c | 13 +-
.../testing/selftests/kvm/x86_64/debug_regs.c | 2 +-
.../kvm/x86_64/userspace_msr_exit_test.c | 9 +-
11 files changed, 195 insertions(+), 37 deletions(-)
--
2.21.3