- Tweak linker script
Ensure that all read only sections end up in one segment, and
everything else in other segments. Discard .eh_frame, .hash and
.gnu.hash since they are unused.
- Don't create invalid string slices in stdout/stderr on Solana
- Report exceeded stack size as a warning if dynamic frames are off
- Native support for signed division in SBF
Adds BPF_SDIV, which is enabled only for the SBF subtarget.
- Introduce dynamic stack frames and the SBFv2 flag
Dynamic stack frames are currently opt-in and enabled setting
cpu=sbfv2. When sbfv2 is used, ELF files are flagged with
e_flags=EF_SBF_V2 so the runtime can detect it and react
accordingly.
(cherry picked from commit 6b611e1c52)
Co-authored-by: Dmitri Makarov <dmakarov@alumni.stanford.edu>
* cargo-build-bpf: don't set -C linker on windows
Since we're now linking using rust-lld
(87ba5c61a5)
which is guaranteed to always be in the sysroot, hardcoding the linker path
shouldn't be needed anymore.
* Update Cargo.lock
- Fix C standard library to include only reentrant versions of functions
- Use a version of solana_rbpf with fixed relocations handling
- Add a test of using C standard library functions in on-chain program
In many Linux distros such as NixOS, the directory in which packages are
installed is assumed to be read-only. To work around this, it is
expected that the filepaths which packaged CLI tools take in are able to
be freely configured through either 1) command-line flags, 2)
environment variables, or 3) a configuration file.
In this commit, environment variables 'BPF_SDK_PATH' and 'BPF_OUT_PATH',
which map respectively to command-line flags '--bpf-sdk-path' and
'--bpf-out-dir', are now handled in cargo-build-bpf.
Additionally, given that arbitrary filepaths may now be set in which the
BPF SDK is located, the requirement in which
'$BPF_SDK_PATH/dependencies/bpf-tools' must strictly be a symbolic link
to the directory '$HOME/.cache/solana/${bpf-tools.version}/bpf-tools has
been relaxed.
Ideally, the directory in which bpf-tools is expected to be downloaded
to and stored should be configurable. Though, this commit serves as a
temporary fix which enables NixOS users to now start being able to build
applications with the Solana SDK.
* Generate a script to rerun the failed sub-command in cargo-test-bpf
* Generate a script to rerun the failed sub-command in cargo-build-bpf
* Add cargo-build-bpf test for generate-child-script-on-failure option
Users need to be able to pass additional command line options such as
-p that are not supported by 'cargo build-bpf' but are meaningful for
the 'cargo build' command. This change allows cargo-build-bpf to pass
all command line options specified after '--' option to 'cargo build'.
rustc default compiler optimization level is O3. This change removes
the option that overrides the default optimization level, because it
is safe to do so. The code generation is incorrect in some cases
because of link-time optimizations, which remain disabled for
compiling BPF programs. In addition, this commit updates the expected
instruction counts for assert_instruction_count test.
Convert to use type alias for the callback and cascade the changes to callers. Thanks @jeffwashington for the help making it possible.
Changed the closure for the progress update in the validator main to FnMut and modify the abort count in the closure which is more reliable.
1. Allow the validator bootstrap code to specify the minimal snapshot download speed. If the snapshot download speed is detected below that, a different RPC can be retried. The default is 10MB/sec.
2. To prevent spinning on a number of sub-optimal choices and not making progress, the abort/retry logic is implemented with the following safe guards:
2.1 at maximum we do this retry for 5 times -- this number is configurable with default 5.
2.2 if the download in one notification round (5 second) is more than 2%, do not do retry -- it is not as bad anyway.
2.3 if the remaining estimate time is less than 1 minutes, do not abort retry as it will be done quickly anyway.
2.4 We do this abort/retry logic only at the first notification to avoid wasting download efforts -- the reasoning is being opportunistic and too greedy may not achieve overall shorter download time.
3. The download_snapshot and download_file is modified with the option allowing caller to notified of download progress via a callback. This allows the business logic of retrying to the place it belongs.
