Protection problems with C++ interfaces

[krlmlr]

My package "bindrcpp" provides a C++ interface via // [[Rcpp::interfaces(cpp)]]. I was able to track down spurious segmentation faults (in dplyr, tidyverse/dplyr#2811) to (what I think is) a protection error in the auto-generated C++ wrappers for these interfaces.

The commit krlmlr/bindrcpp@6d1e2bd shows both the problem and the solution: In general, wrap() will perform allocation of R objects, and without extra shielding these can be garbage collected even before the wrapped function is called, because the latter has all SEXP in its signature.

After applying this particular commit, I'm not seeing test failures anymore, which before occurred after a few iterations. Happy to submit a PR.

CC @lionel-.

[eddelbuettel]

It's been a while but we did revise that aspect once or twice already.

Could you provide a reproducible example we could call before and after? With and without gctorture()?

[krlmlr]

Thanks. That commit is the reprex, gctorture2() is called inside the tests:

hub clone krlmlr/bindrcpp
cd bindrcpp
git checkout 6d1e2bd^ # omit ^ to see fixed behavior
R -q -e 'devtools::test(reporter = c("summary", "stop"))'

[eddelbuettel]

I am sorry but I don't use hub, or devtools.

And involving an entire different package strikes me as a little far away from "minimal".

[krlmlr]

Give me a minute to translate that for you. I seriously recommend checking out hub, though: https://github.com/github/hub

[eddelbuettel]

Yes, I read about it. So far exiting command-line and magit are good enough for me.

[lionel-]

who needs hub when there's magit ;)

[krlmlr]

Thanks for the pointers, will check out magit.

Now, for the reprex:

1. Install bindrcpp from https://github.com/krlmlr/bindrcpp/tree/a7759861e780fa84cf022f31809af3744599abea
2. Run the script below, it should stop very soon
3. Now install the fixed version from https://github.com/krlmlr/bindrcpp/tree/6d1e2bde24fd8a26a3911923cd2e6cd0a1a29180
4. Run the script below, it should run until the end

If you install from a clone with R CMD INSTALL ., you need to remove src/*.o before switching to a different version.

library(bindrcpp)
gctorture2(29)

for (i in 1:1000) {

print(i)

env_cb <- bindrcpp:::cpp_create_environment(letters, "toupper")
env <- env_cb$env
env2_cb <- bindrcpp:::cpp_create_environment(LETTERS, "tolower", parent = env)
env2 <- env2_cb$env
stopifnot(identical(get("a", env2), "A"))
stopifnot(identical(get("x", env2), "X"))
stopifnot(identical(env2$a, NULL))
stopifnot(identical(env2$x, NULL))
stopifnot(identical(env2$B, "b"))
stopifnot(identical(env2$Y, "y"))
stopifnot(identical(length(ls(env2)), length(letters)))
env2$a <- "a"
stopifnot(identical(get("a", env2), "a"))

}

[jjallaire]

PR is here: #713

…

On Wed, Jun 14, 2017 at 8:45 AM, Kirill Müller ***@***.***> wrote: Thanks for the pointers, will check out magit. Now, for the reprex: 1. Install bindrcpp from https://github.com/krlmlr/bindrcpp/tree/ a7759861e780fa84cf022f31809af3744599abea <https://github.com/krlmlr/bindrcpp/tree/a7759861e780fa84cf022f31809af3744599abea> 2. Run the script below, it should stop very soon 3. Now install the fixed version from https://github.com/krlmlr/ bindrcpp/tree/6d1e2bde24fd8a26a3911923cd2e6cd0a1a29180 <https://github.com/krlmlr/bindrcpp/tree/6d1e2bde24fd8a26a3911923cd2e6cd0a1a29180> 4. Run the script below, it should run until the end If you install from a clone with R CMD INSTALL ., you need to remove src/*.o before switching to a different version. library(bindrcpp) gctorture2(29) for (i in 1:1000) { print(i) env_cb <- bindrcpp:::cpp_create_environment(letters, "toupper")env <- env_cb$envenv2_cb <- bindrcpp:::cpp_create_environment(LETTERS, "tolower", parent = env)env2 <- env2_cb$env stopifnot(identical(get("a", env2), "A")) stopifnot(identical(get("x", env2), "X")) stopifnot(identical(env2$a, NULL)) stopifnot(identical(env2$x, NULL)) stopifnot(identical(env2$B, "b")) stopifnot(identical(env2$Y, "y")) stopifnot(identical(length(ls(env2)), length(letters)))#expect_error(env2$B <- "B", "read-only")env2$a <- "a" stopifnot(identical(get("a", env2), "a")) } — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#712 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAGXx0z3KK_9S8UeYOTrwAMya8ICCDlNks5sD9XWgaJpZM4N5zBG> .