diff --git a/adb/daemon/shell_service.cpp b/adb/daemon/shell_service.cpp
index 0fb14c4..f62032d 100644
--- a/adb/daemon/shell_service.cpp
+++ b/adb/daemon/shell_service.cpp
@@ -85,7 +85,6 @@
 #include <paths.h>
 #include <pty.h>
 #include <pwd.h>
-#include <sys/select.h>
 #include <termios.h>
 
 #include <memory>
@@ -141,6 +140,20 @@
     return true;
 }
 
+struct SubprocessPollfds {
+    adb_pollfd pfds[3];
+
+    adb_pollfd* data() { return pfds; }
+    size_t size() { return 3; }
+
+    adb_pollfd* begin() { return pfds; }
+    adb_pollfd* end() { return pfds + size(); }
+
+    adb_pollfd& stdinout_pfd() { return pfds[0]; }
+    adb_pollfd& stderr_pfd() { return pfds[1]; }
+    adb_pollfd& protocol_pfd() { return pfds[2]; }
+};
+
 class Subprocess {
   public:
     Subprocess(std::string command, const char* terminal_type, SubprocessType type,
@@ -176,8 +189,7 @@
     void PassDataStreams();
     void WaitForExit();
 
-    unique_fd* SelectLoop(fd_set* master_read_set_ptr,
-                          fd_set* master_write_set_ptr);
+    unique_fd* PollLoop(SubprocessPollfds* pfds);
 
     // Input/output stream handlers. Success returns nullptr, failure returns
     // a pointer to the failed FD.
@@ -545,23 +557,23 @@
     }
 
     // Start by trying to read from the protocol FD, stdout, and stderr.
-    fd_set master_read_set, master_write_set;
-    FD_ZERO(&master_read_set);
-    FD_ZERO(&master_write_set);
-    for (unique_fd* sfd : {&protocol_sfd_, &stdinout_sfd_, &stderr_sfd_}) {
-        if (*sfd != -1) {
-            FD_SET(sfd->get(), &master_read_set);
-        }
-    }
+    SubprocessPollfds pfds;
+    pfds.stdinout_pfd() = {.fd = stdinout_sfd_.get(), .events = POLLIN};
+    pfds.stderr_pfd() = {.fd = stderr_sfd_.get(), .events = POLLIN};
+    pfds.protocol_pfd() = {.fd = protocol_sfd_.get(), .events = POLLIN};
 
     // Pass data until the protocol FD or both the subprocess pipes die, at
     // which point we can't pass any more data.
     while (protocol_sfd_ != -1 && (stdinout_sfd_ != -1 || stderr_sfd_ != -1)) {
-        unique_fd* dead_sfd = SelectLoop(&master_read_set, &master_write_set);
+        unique_fd* dead_sfd = PollLoop(&pfds);
         if (dead_sfd) {
             D("closing FD %d", dead_sfd->get());
-            FD_CLR(dead_sfd->get(), &master_read_set);
-            FD_CLR(dead_sfd->get(), &master_write_set);
+            auto it = std::find_if(pfds.begin(), pfds.end(), [=](const adb_pollfd& pfd) {
+                return pfd.fd == dead_sfd->get();
+            });
+            CHECK(it != pfds.end());
+            it->fd = -1;
+            it->events = 0;
             if (dead_sfd == &protocol_sfd_) {
                 // Using SIGHUP is a decent general way to indicate that the
                 // controlling process is going away. If specific signals are
@@ -583,30 +595,19 @@
     }
 }
 
-namespace {
-
-inline bool ValidAndInSet(const unique_fd& sfd, fd_set* set) {
-    return sfd != -1 && FD_ISSET(sfd.get(), set);
-}
-
-}   // namespace
-
-unique_fd* Subprocess::SelectLoop(fd_set* master_read_set_ptr,
-                                  fd_set* master_write_set_ptr) {
-    fd_set read_set, write_set;
-    int select_n =
-            std::max(std::max(protocol_sfd_.get(), stdinout_sfd_.get()), stderr_sfd_.get()) + 1;
+unique_fd* Subprocess::PollLoop(SubprocessPollfds* pfds) {
     unique_fd* dead_sfd = nullptr;
+    adb_pollfd& stdinout_pfd = pfds->stdinout_pfd();
+    adb_pollfd& stderr_pfd = pfds->stderr_pfd();
+    adb_pollfd& protocol_pfd = pfds->protocol_pfd();
 
-    // Keep calling select() and passing data until an FD closes/errors.
+    // Keep calling poll() and passing data until an FD closes/errors.
     while (!dead_sfd) {
-        memcpy(&read_set, master_read_set_ptr, sizeof(read_set));
-        memcpy(&write_set, master_write_set_ptr, sizeof(write_set));
-        if (select(select_n, &read_set, &write_set, nullptr, nullptr) < 0) {
+        if (adb_poll(pfds->data(), pfds->size(), -1) < 0) {
             if (errno == EINTR) {
                 continue;
             } else {
-                PLOG(ERROR) << "select failed, closing subprocess pipes";
+                PLOG(ERROR) << "poll failed, closing subprocess pipes";
                 stdinout_sfd_.reset(-1);
                 stderr_sfd_.reset(-1);
                 return nullptr;
@@ -614,34 +615,47 @@
         }
 
         // Read stdout, write to protocol FD.
-        if (ValidAndInSet(stdinout_sfd_, &read_set)) {
+        if (stdinout_pfd.fd != -1 && (stdinout_pfd.revents & POLLIN)) {
             dead_sfd = PassOutput(&stdinout_sfd_, ShellProtocol::kIdStdout);
         }
 
         // Read stderr, write to protocol FD.
-        if (!dead_sfd && ValidAndInSet(stderr_sfd_, &read_set)) {
+        if (!dead_sfd && stderr_pfd.fd != 1 && (stderr_pfd.revents & POLLIN)) {
             dead_sfd = PassOutput(&stderr_sfd_, ShellProtocol::kIdStderr);
         }
 
         // Read protocol FD, write to stdin.
-        if (!dead_sfd && ValidAndInSet(protocol_sfd_, &read_set)) {
+        if (!dead_sfd && protocol_pfd.fd != -1 && (protocol_pfd.revents & POLLIN)) {
             dead_sfd = PassInput();
             // If we didn't finish writing, block on stdin write.
             if (input_bytes_left_) {
-                FD_CLR(protocol_sfd_.get(), master_read_set_ptr);
-                FD_SET(stdinout_sfd_.get(), master_write_set_ptr);
+                protocol_pfd.events &= ~POLLIN;
+                stdinout_pfd.events |= POLLOUT;
             }
         }
 
         // Continue writing to stdin; only happens if a previous write blocked.
-        if (!dead_sfd && ValidAndInSet(stdinout_sfd_, &write_set)) {
+        if (!dead_sfd && stdinout_pfd.fd != -1 && (stdinout_pfd.revents & POLLOUT)) {
             dead_sfd = PassInput();
             // If we finished writing, go back to blocking on protocol read.
             if (!input_bytes_left_) {
-                FD_SET(protocol_sfd_.get(), master_read_set_ptr);
-                FD_CLR(stdinout_sfd_.get(), master_write_set_ptr);
+                protocol_pfd.events |= POLLIN;
+                stdinout_pfd.events &= ~POLLOUT;
             }
         }
+
+        // After handling all of the events we've received, check to see if any fds have died.
+        if (stdinout_pfd.revents & (POLLHUP | POLLRDHUP | POLLERR | POLLNVAL)) {
+            return &stdinout_sfd_;
+        }
+
+        if (stderr_pfd.revents & (POLLHUP | POLLRDHUP | POLLERR | POLLNVAL)) {
+            return &stderr_sfd_;
+        }
+
+        if (protocol_pfd.revents & (POLLHUP | POLLRDHUP | POLLERR | POLLNVAL)) {
+            return &protocol_sfd_;
+        }
     }  // while (!dead_sfd)
 
     return dead_sfd;
diff --git a/adb/test_device.py b/adb/test_device.py
index 57925e8..083adce 100755
--- a/adb/test_device.py
+++ b/adb/test_device.py
@@ -536,6 +536,36 @@
         for i, success in result.iteritems():
             self.assertTrue(success)
 
+    def disabled_test_parallel(self):
+        """Spawn a bunch of `adb shell` instances in parallel.
+
+        This was broken historically due to the use of select, which only works
+        for fds that are numerically less than 1024.
+
+        Bug: http://b/141955761"""
+
+        n_procs = 2048
+        procs = dict()
+        for i in xrange(0, n_procs):
+            procs[i] = subprocess.Popen(
+                ['adb', 'shell', 'read foo; echo $foo; read rc; exit $rc'],
+                stdin=subprocess.PIPE,
+                stdout=subprocess.PIPE
+            )
+
+        for i in xrange(0, n_procs):
+            procs[i].stdin.write("%d\n" % i)
+
+        for i in xrange(0, n_procs):
+            response = procs[i].stdout.readline()
+            assert(response == "%d\n" % i)
+
+        for i in xrange(0, n_procs):
+            procs[i].stdin.write("%d\n" % (i % 256))
+
+        for i in xrange(0, n_procs):
+            assert(procs[i].wait() == i % 256)
+
 
 class ArgumentEscapingTest(DeviceTest):
     def test_shell_escaping(self):