diff --git a/libc/bionic/sysconf.cpp b/libc/bionic/sysconf.cpp
index 2b3200c..dd6b129 100644
--- a/libc/bionic/sysconf.cpp
+++ b/libc/bionic/sysconf.cpp
@@ -52,16 +52,31 @@
     // Things we actually have to calculate...
     //
     case _SC_ARG_MAX:
-      // https://lkml.org/lkml/2017/11/15/813...
+      // You might think that just returning a constant 128KiB (ARG_MAX) would
+      // make sense, as this guy did:
       //
-      // I suspect a 128kB sysconf(_SC_ARG_MAX) is the sanest bet, simply
-      // because of that "conservative is better than aggressive".
+      //   https://lkml.org/lkml/2017/11/15/813...
       //
-      // Especially since _technically_ we're still limiting things to that
-      // 128kB due to the single-string limit.
+      //   I suspect a 128kB sysconf(_SC_ARG_MAX) is the sanest bet, simply
+      //   because of that "conservative is better than aggressive".
       //
-      //               Linus
-      return ARG_MAX;
+      //   Especially since _technically_ we're still limiting things to that
+      //   128kB due to the single-string limit.
+      //
+      //                 Linus
+      //
+      // In practice that caused us trouble with toybox tests for xargs edge
+      // cases. The tests assume that they can at least reach the kernel's
+      // "minimum maximum" of 128KiB, but if we report 128KiB for _SC_ARG_MAX
+      // and xargs starts subtracting the environment space and so on from that,
+      // then xargs will think it's run out of space when given 128KiB of data,
+      // which should always work. See this thread for more:
+      //
+      // http://lists.landley.net/pipermail/toybox-landley.net/2019-November/011229.html
+      //
+      // So let's resign ourselves to tracking what the kernel actually does.
+      // Right now (2019, Linux 5.3) that amounts to:
+      return MAX(MIN(__sysconf_rlimit(RLIMIT_STACK) / 4, 3 * _STK_LIM / 4), ARG_MAX);
 
     case _SC_AVPHYS_PAGES:      return get_avphys_pages();
     case _SC_CHILD_MAX:         return __sysconf_rlimit(RLIMIT_NPROC);
diff --git a/tests/unistd_test.cpp b/tests/unistd_test.cpp
index 99d92e6..f3b08c3 100644
--- a/tests/unistd_test.cpp
+++ b/tests/unistd_test.cpp
@@ -27,6 +27,7 @@
 #include <stdint.h>
 #include <sys/capability.h>
 #include <sys/param.h>
+#include <sys/resource.h>
 #include <sys/syscall.h>
 #include <sys/types.h>
 #include <sys/utsname.h>
@@ -1065,11 +1066,38 @@
 }
 
 TEST(UNISTD_TEST, sysconf_SC_ARG_MAX) {
-  // https://lkml.org/lkml/2017/11/15/813.
-#if !defined(ARG_MAX)
-#define ARG_MAX 131072
-#endif
-  ASSERT_EQ(ARG_MAX, sysconf(_SC_ARG_MAX));
+  // Since Linux 2.6.23, ARG_MAX isn't a constant and depends on RLIMIT_STACK.
+  // See prepare_arg_pages() in the kernel for the gory details:
+  // https://elixir.bootlin.com/linux/v5.3.11/source/fs/exec.c#L451
+
+  // Get our current limit, and set things up so we restore the limit.
+  rlimit rl;
+  ASSERT_EQ(0, getrlimit(RLIMIT_STACK, &rl));
+  uint64_t original_rlim_cur = rl.rlim_cur;
+  if (rl.rlim_cur == RLIM_INFINITY) {
+    rl.rlim_cur = 8 * 1024 * 1024; // Bionic reports unlimited stacks as 8MiB.
+  }
+  auto guard = android::base::make_scope_guard([&rl, original_rlim_cur]() {
+    rl.rlim_cur = original_rlim_cur;
+    ASSERT_EQ(0, setrlimit(RLIMIT_STACK, &rl));
+  });
+
+  // _SC_ARG_MAX should be 1/4 the stack size.
+  EXPECT_EQ(static_cast<long>(rl.rlim_cur / 4), sysconf(_SC_ARG_MAX));
+
+  // If you have a really small stack, the kernel still guarantees "32 pages" (see fs/exec.c).
+  rl.rlim_cur = 1024;
+  rl.rlim_max = RLIM_INFINITY;
+  ASSERT_EQ(0, setrlimit(RLIMIT_STACK, &rl));
+
+  EXPECT_EQ(static_cast<long>(32 * sysconf(_SC_PAGE_SIZE)), sysconf(_SC_ARG_MAX));
+
+  // With a 128-page stack limit, we know exactly what _SC_ARG_MAX should be...
+  rl.rlim_cur = 128 * sysconf(_SC_PAGE_SIZE);
+  rl.rlim_max = RLIM_INFINITY;
+  ASSERT_EQ(0, setrlimit(RLIMIT_STACK, &rl));
+
+  EXPECT_EQ(static_cast<long>((128 * sysconf(_SC_PAGE_SIZE)) / 4), sysconf(_SC_ARG_MAX));
 }
 
 TEST(UNISTD_TEST, sysconf_unknown) {