The TPU userspace driver needs access to specific PCI device information
located in Linux sysfs. We mirror the sysfs paths the driver reads on the host
in the Sentry sysfs. This way we can ensure we only expose the host device
information that's strictly necessary for TPU to run.
PiperOrigin-RevId: 550005271
Tested on 1 V100 GPU:
```
$ docker run --runtime=runsc --rm --gpus all nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda11.7.1-ubi8
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED
Done
```
PiperOrigin-RevId: 549837326
Update the memmap IncRef method to pass memory cgroup id and store it in the
FrameRefSet which will be used for memory accounting. During DecRef, the
memCgID from the FrameRefSet will be retrieved and passed to MemoryLocked.Dec
to remove the memory from the cgroup.
PiperOrigin-RevId: 549656411
We were only missing `AMPERE_COMPUTE_B` allocation class.
Verified on an AWS g5.xlarge instance that runsc can at least run basic
vectoradd cuda application.
```
$ docker run --runtime=runsc --rm --gpus all --runtime=runsc nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda11.7.1-ubi8
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED
Done
```
Fixes#9151
PiperOrigin-RevId: 547560783
`git diff` shows no relevant changes in
`src/common/sdk/nvidia/inc/{alloc,class,ctrl}/`,
`src/common/sdk/nvidia/inc/nvos.h`, `kernel-open/common/inc/`, or
`kernel-open/nvidia-uvm/uvm_ioctl.h`. Manually tested with T4, A100, L4.
PiperOrigin-RevId: 540992697
With this change, we can now run simple CUDA applications on H100 GPUs.
```
$ docker run --runtime=runsc --gpus all nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda11.7.1-ubi8
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED
Done
```
Note that this was tested with 525.60.13 driver version.
Updates #14
PiperOrigin-RevId: 534607359
This catches up the interface to the `EmitUnimplementedEvent` method signature
on `kernel.Kernel`.
Also add build-time test to verify that `kernel.Kernel` implements this
interface, in order to catch such breakages at build time in the future.
PiperOrigin-RevId: 519000411
The tun code does not properly abide by reference counting rules. Change it
so the ownership is more clear and to prevent potential use-after-free.
PiperOrigin-RevId: 482865862
This change has significant performance implications. bufferv2 is reference
counted and pooled, which alleviates heap/GC pressure. Below are the results
from running the iperf benchmark.
HEAD:
BenchmarkIperf/operation.Upload-16 1552 ns/op 46.6GiB total allocations
BenchmarkIperf/operation.Download-16 1114 ns/op 68.6GiB total allocations
w/ change:
BenchmarkIperf/operation.Upload-16 1139 ns/op (-27%) 1.41GiB total allocations (-97%)
BenchmarkIperf/operation.Download-16 753.2 ns/op (-33%) 706MiB total allocations (-99%)
PiperOrigin-RevId: 462453185
All atomic 64 bit ints are changed to atomicbitops.(Ui|I)nt64. A nogo checker
enforces that sync/atomic 64 bit functions are not called.
For reviewers: the interesting changes are in the atomicbitops and checkaligned
packages.
Why do this?
- It is very easy to accidentally use atomic values without sync/atomic funcs.
- We have checkatomics, but this is optional and is forgotten in several places.
- Using a type+checker to enforce this seems less error prone and simpler.
- We get NoCopy protection.
- Use of 64 bit atomics can break 32 bit builds. We have types to handle this
without any runtime cost, so we might as well use them.
PiperOrigin-RevId: 440473398
Linux behaves differently for regular files and dirs for poll(2)/select(2)
compared to epoll_ctl(2). The latter returns EPERM for file and dirs.
I've also changed host FDs to behave like the underlying FD in regards
to epoll to keep it compatible with docker.
Fixes#7134
PiperOrigin-RevId: 429412692
Docker maps stdin to `/dev/null` which doesn't support epoll. Host FD
was ignoring the error and suceeding the epoll_ctl call from the
container, giving false impressing that epoll would be notified.
This required plumbing failure to all waiter.Waitable.EventRegister
callers and implementers.
Closes#6795
PiperOrigin-RevId: 414797621
Instead of passing the event mask at registratrion time, pass the mask as part
of the waiter. This makes the mask immutable and simplifies the architecture of
waiters. This is also necessary for a future fix that will allow the fdnotifier
to keep persistent entries, as opposed to requiring constant updates.
This change is intended to be a no-op in terms of function. The only exception
is signalfd, where this mask was abused. To handle this case, the operation of
signalfd changed to allow one layer of indirection.
PiperOrigin-RevId: 409702998
Update all instances of the above errors to the faster linuxerr implementation.
With the temporary linuxerr.Equals(), no logical changes are made.
PiperOrigin-RevId: 382306655
Remove three syserror entries duplicated in linuxerr. Because of the
linuxerr.Equals method, this is a mere change of return values from
syserror to linuxerr definitions.
Done with only these three errnos as CLs removing all grow to a significantly
large size.
PiperOrigin-RevId: 382173835
Split usermem package to help remove syserror dependency in go_marshal.
New hostarch package contains code not dependent on syserror.
PiperOrigin-RevId: 365651233
Doing so involved breaking dependencies between //pkg/tcpip and the rest
of gVisor, which are discouraged anyways.
Tested on the Go branch via:
gvisor.dev/gvisor/pkg/tcpip/...
Addresses #1446.
PiperOrigin-RevId: 363081778
- Make AddressableEndpoint optional for NetworkEndpoint.
Not all NetworkEndpoints need to support addressing (e.g. ARP), so
AddressableEndpoint should only be implemented for protocols that
support addressing such as IPv4 and IPv6.
With this change, tcpip.ErrNotSupported will be returned by the stack
when attempting to modify addresses on a network endpoint that does
not support addressing.
Now that packets are fully handled at the network layer, and (with this
change) addresses are optional for network endpoints, we no longer need
the workaround for ARP where a fake ARP address was added to each NIC
that performs ARP so that packets would be delivered to the ARP layer.
PiperOrigin-RevId: 342722547
Inode number consistency checks are now skipped in save/restore tests for
reasons described in greatest detail in StatTest.StateDoesntChangeAfterRename.
They pass in VFS1 due to the bug described in new test case
SimpleStatTest.DifferentFilesHaveDifferentDeviceInodeNumberPairs.
Fixes#1663
PiperOrigin-RevId: 338776148
When a response needs to be sent to an incoming packet, the stack should
consult its neighbour table to determine the remote address's link
address.
When an entry does not exist in the stack's neighbor table, the stack
should queue the packet while link resolution completes. See comments.
PiperOrigin-RevId: 336185457
This is more consistent with Linux (see comment on MM.NewSharedAnonMappable()).
We don't do the same thing on VFS1 for reasons documented by the updated
comment.
PiperOrigin-RevId: 332514849
This is in compliance with VFS1. See pkg/sentry/fs/dev/tty.go in the struct
ttyInodeOperations.
Fixes the failure of python runtime test_ioctl.
Updates #3515
PiperOrigin-RevId: 327042758
LockFD is the generic implementation that can be embedded in
FileDescriptionImpl implementations. Unique lock ID is
maintained in vfs.FileDescription and is created on demand.
Updates #1480
PiperOrigin-RevId: 315604825
