diff --git a/debuggerd/debuggerd_test.cpp b/debuggerd/debuggerd_test.cpp
index 64df53e..1f0e420 100644
--- a/debuggerd/debuggerd_test.cpp
+++ b/debuggerd/debuggerd_test.cpp
@@ -243,9 +243,10 @@
 
 void CrasherTest::AssertDeath(int signo) {
   int status;
-  pid_t pid = TIMEOUT(5, waitpid(crasher_pid, &status, 0));
+  pid_t pid = TIMEOUT(10, waitpid(crasher_pid, &status, 0));
   if (pid != crasher_pid) {
-    printf("failed to wait for crasher (pid %d)\n", crasher_pid);
+    printf("failed to wait for crasher (expected pid %d, return value %d): %s\n", crasher_pid, pid,
+           strerror(errno));
     sleep(100);
     FAIL() << "failed to wait for crasher: " << strerror(errno);
   }
diff --git a/libmeminfo/libdmabufinfo/tools/dmabuf_dump.cpp b/libmeminfo/libdmabufinfo/tools/dmabuf_dump.cpp
index 0851fb3..9a80c82 100644
--- a/libmeminfo/libdmabufinfo/tools/dmabuf_dump.cpp
+++ b/libmeminfo/libdmabufinfo/tools/dmabuf_dump.cpp
@@ -17,16 +17,17 @@
 #include <dirent.h>
 #include <errno.h>
 #include <inttypes.h>
+#include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 
-#include <iostream>
 #include <fstream>
+#include <iostream>
+#include <map>
+#include <set>
 #include <sstream>
 #include <string>
 #include <vector>
-#include <map>
-#include <set>
 
 #include <android-base/stringprintf.h>
 #include <dmabufinfo/dmabufinfo.h>
@@ -43,7 +44,7 @@
     exit(exit_status);
 }
 
-static std::string GetProcessBaseName(pid_t pid) {
+static std::string GetProcessComm(const pid_t pid) {
     std::string pid_path = android::base::StringPrintf("/proc/%d/comm", pid);
     std::ifstream in{pid_path};
     if (!in) return std::string("N/A");
@@ -53,85 +54,76 @@
     return line;
 }
 
-static void AddPidsToSet(const std::unordered_map<pid_t, int>& map, std::set<pid_t>* set)
-{
-    for (auto it = map.begin(); it != map.end(); ++it)
-        set->insert(it->first);
+static void AddPidsToSet(const std::unordered_map<pid_t, int>& map, std::set<pid_t>* set) {
+    for (auto it = map.begin(); it != map.end(); ++it) set->insert(it->first);
 }
 
 static void PrintDmaBufInfo(const std::vector<DmaBuffer>& bufs) {
-    std::set<pid_t> pid_set;
-    std::map<pid_t, int> pid_column;
-
     if (bufs.empty()) {
-        std::cout << "dmabuf info not found ¯\\_(ツ)_/¯" << std::endl;
+        printf("dmabuf info not found ¯\\_(ツ)_/¯\n");
         return;
     }
 
     // Find all unique pids in the input vector, create a set
+    std::set<pid_t> pid_set;
     for (int i = 0; i < bufs.size(); i++) {
         AddPidsToSet(bufs[i].fdrefs(), &pid_set);
         AddPidsToSet(bufs[i].maprefs(), &pid_set);
     }
 
-    int pid_count = 0;
-
-    std::cout << "\t\t\t\t\t\t";
-
-    // Create a map to convert each unique pid into a column number
-    for (auto it = pid_set.begin(); it != pid_set.end(); ++it, ++pid_count) {
-        pid_column.insert(std::make_pair(*it, pid_count));
-        std::cout << ::android::base::StringPrintf("[pid: % 4d]\t", *it);
+    // Format the header string spaced and separated with '|'
+    printf("    Dmabuf Inode |            Size |      Ref Counts |");
+    for (auto pid : pid_set) {
+        printf("%16s:%-5d |", GetProcessComm(pid).c_str(), pid);
     }
+    printf("\n");
 
-    std::cout << std::endl << "\t\t\t\t\t\t";
+    // holds per-process dmabuf size in kB
+    std::map<pid_t, uint64_t> per_pid_size = {};
+    uint64_t dmabuf_total_size = 0;
 
-    for (auto it = pid_set.begin(); it != pid_set.end(); ++it) {
-        std::cout << ::android::base::StringPrintf("%16s",
-            GetProcessBaseName(*it).c_str());
-    }
+    // Iterate through all dmabufs and collect per-process sizes, refs
+    for (auto& buf : bufs) {
+        printf("%16ju |%13" PRIu64 " kB |%16" PRIu64 " |", static_cast<uintmax_t>(buf.inode()),
+               buf.size() / 1024, buf.total_refs());
+        // Iterate through each process to find out per-process references for each buffer,
+        // gather total size used by each process etc.
+        for (pid_t pid : pid_set) {
+            int pid_refs = 0;
+            if (buf.fdrefs().count(pid) == 1) {
+                // Get the total number of ref counts the process is holding
+                // on this buffer. We don't differentiate between mmap or fd.
+                pid_refs += buf.fdrefs().at(pid);
+                if (buf.maprefs().count(pid) == 1) {
+                    pid_refs += buf.maprefs().at(pid);
+                }
+            }
 
-    std::cout << std::endl << "\tinode\t\tsize\t\tcount\t";
-    for (int i = 0; i < pid_count; i++) {
-        std::cout << "fd\tmap\t";
-    }
-    std::cout << std::endl;
-
-    auto fds = std::make_unique<int[]>(pid_count);
-    auto maps = std::make_unique<int[]>(pid_count);
-    auto pss = std::make_unique<long[]>(pid_count);
-
-    memset(pss.get(), 0, sizeof(long) * pid_count);
-
-    for (auto buf = bufs.begin(); buf != bufs.end(); ++buf) {
-
-        std::cout << ::android::base::StringPrintf("%16lu\t%10" PRIu64 "\t%" PRIu64 "\t",
-            buf->inode(),buf->size(), buf->count());
-
-        memset(fds.get(), 0, sizeof(int) * pid_count);
-        memset(maps.get(), 0, sizeof(int) * pid_count);
-
-        for (auto it = buf->fdrefs().begin(); it != buf->fdrefs().end(); ++it) {
-            fds[pid_column[it->first]] = it->second;
-            pss[pid_column[it->first]] += buf->size() * it->second / buf->count();
+            if (pid_refs) {
+                // Add up the per-pid total size. Note that if a buffer is mapped
+                // in 2 different processes, the size will be shown as mapped or opened
+                // in both processes. This is intended for visibility.
+                //
+                // If one wants to get the total *unique* dma buffers, they can simply
+                // sum the size of all dma bufs shown by the tool
+                per_pid_size[pid] += buf.size() / 1024;
+                printf("%17d refs |", pid_refs);
+            } else {
+                printf("%22s |", "--");
+            }
         }
-
-        for (auto it = buf->maprefs().begin(); it != buf->maprefs().end(); ++it) {
-            maps[pid_column[it->first]] = it->second;
-            pss[pid_column[it->first]] += buf->size() * it->second / buf->count();
-        }
-
-        for (int i = 0; i < pid_count; i++) {
-            std::cout << ::android::base::StringPrintf("%d\t%d\t", fds[i], maps[i]);
-        }
-        std::cout << std::endl;
+        dmabuf_total_size += buf.size() / 1024;
+        printf("\n");
     }
-    std::cout << "-----------------------------------------" << std::endl;
-    std::cout << "PSS                                      ";
-    for (int i = 0; i < pid_count; i++) {
-        std::cout << ::android::base::StringPrintf("%15ldK", pss[i] / 1024);
+
+    printf("------------------------------------\n");
+    printf("%-16s  %13" PRIu64 " kB |%16s |", "TOTALS", dmabuf_total_size, "n/a");
+    for (auto pid : pid_set) {
+        printf("%19" PRIu64 " kB |", per_pid_size[pid]);
     }
-    std::cout << std::endl;
+    printf("\n");
+
+    return;
 }
 
 int main(int argc, char* argv[]) {
@@ -142,8 +134,7 @@
     if (argc > 1) {
         if (sscanf(argv[1], "%d", &pid) == 1) {
             show_all = false;
-        }
-        else {
+        } else {
             usage(EXIT_FAILURE);
         }
     }
@@ -178,8 +169,7 @@
             exit(EXIT_FAILURE);
         }
     }
+
     PrintDmaBufInfo(bufs);
     return 0;
 }
-
-
diff --git a/libmeminfo/pageacct.cpp b/libmeminfo/pageacct.cpp
index 0a26c08..cb17af8 100644
--- a/libmeminfo/pageacct.cpp
+++ b/libmeminfo/pageacct.cpp
@@ -81,7 +81,8 @@
         if (!InitPageAcct()) return false;
     }
 
-    if (pread64(kpageflags_fd_, flags, sizeof(uint64_t), pfn * sizeof(uint64_t)) < 0) {
+    if (pread64(kpageflags_fd_, flags, sizeof(uint64_t), pfn * sizeof(uint64_t)) !=
+        sizeof(uint64_t)) {
         PLOG(ERROR) << "Failed to read page flags for page " << pfn;
         return false;
     }
@@ -95,7 +96,8 @@
         if (!InitPageAcct()) return false;
     }
 
-    if (pread64(kpagecount_fd_, mapcount, sizeof(uint64_t), pfn * sizeof(uint64_t)) < 0) {
+    if (pread64(kpagecount_fd_, mapcount, sizeof(uint64_t), pfn * sizeof(uint64_t)) !=
+        sizeof(uint64_t)) {
         PLOG(ERROR) << "Failed to read map count for page " << pfn;
         return false;
     }
@@ -130,7 +132,7 @@
     off64_t offset = pfn_to_idle_bitmap_offset(pfn);
     uint64_t idle_bits;
 
-    if (pread64(pageidle_fd_, &idle_bits, sizeof(uint64_t), offset) < 0) {
+    if (pread64(pageidle_fd_, &idle_bits, sizeof(uint64_t), offset) != sizeof(uint64_t)) {
         PLOG(ERROR) << "Failed to read page idle bitmap for page " << pfn;
         return -errno;
     }
diff --git a/libmeminfo/procmeminfo.cpp b/libmeminfo/procmeminfo.cpp
index 934d65c..caf6e86 100644
--- a/libmeminfo/procmeminfo.cpp
+++ b/libmeminfo/procmeminfo.cpp
@@ -27,6 +27,7 @@
 #include <memory>
 #include <string>
 #include <utility>
+#include <vector>
 
 #include <android-base/file.h>
 #include <android-base/logging.h>
@@ -278,68 +279,89 @@
 
 bool ProcMemInfo::ReadVmaStats(int pagemap_fd, Vma& vma, bool get_wss, bool use_pageidle) {
     PageAcct& pinfo = PageAcct::Instance();
-    uint64_t pagesz = getpagesize();
-    uint64_t num_pages = (vma.end - vma.start) / pagesz;
-
-    std::unique_ptr<uint64_t[]> pg_frames(new uint64_t[num_pages]);
-    uint64_t first = vma.start / pagesz;
-    if (pread64(pagemap_fd, pg_frames.get(), num_pages * sizeof(uint64_t),
-                first * sizeof(uint64_t)) < 0) {
-        PLOG(ERROR) << "Failed to read page frames from page map for pid: " << pid_;
+    if (get_wss && use_pageidle && !pinfo.InitPageAcct(true)) {
+        LOG(ERROR) << "Failed to init idle page accounting";
         return false;
     }
 
-    if (get_wss && use_pageidle) {
-        if (!pinfo.InitPageAcct(true)) {
-            LOG(ERROR) << "Failed to init idle page accounting";
-            return false;
-        }
-    }
+    uint64_t pagesz = getpagesize();
+    size_t num_pages = (vma.end - vma.start) / pagesz;
+    size_t first_page = vma.start / pagesz;
 
-    std::unique_ptr<uint64_t[]> pg_flags(new uint64_t[num_pages]);
-    std::unique_ptr<uint64_t[]> pg_counts(new uint64_t[num_pages]);
-    for (uint64_t i = 0; i < num_pages; ++i) {
+    std::vector<uint64_t> page_cache;
+    size_t cur_page_cache_index = 0;
+    size_t num_in_page_cache = 0;
+    size_t num_leftover_pages = num_pages;
+    for (size_t cur_page = first_page; cur_page < first_page + num_pages; ++cur_page) {
         if (!get_wss) {
             vma.usage.vss += pagesz;
         }
-        uint64_t p = pg_frames[i];
-        if (!PAGE_PRESENT(p) && !PAGE_SWAPPED(p)) continue;
 
-        if (PAGE_SWAPPED(p)) {
+        // Cache page map data.
+        if (cur_page_cache_index == num_in_page_cache) {
+            static constexpr size_t kMaxPages = 2048;
+            num_leftover_pages -= num_in_page_cache;
+            if (num_leftover_pages > kMaxPages) {
+                num_in_page_cache = kMaxPages;
+            } else {
+                num_in_page_cache = num_leftover_pages;
+            }
+            page_cache.resize(num_in_page_cache);
+            size_t total_bytes = page_cache.size() * sizeof(uint64_t);
+            ssize_t bytes = pread64(pagemap_fd, page_cache.data(), total_bytes,
+                                    cur_page * sizeof(uint64_t));
+            if (bytes != total_bytes) {
+                if (bytes == -1) {
+                    PLOG(ERROR) << "Failed to read page data at offset "
+                                << cur_page * sizeof(uint64_t);
+                } else {
+                    LOG(ERROR) << "Failed to read page data at offset "
+                               << cur_page * sizeof(uint64_t) << " read bytes " << sizeof(uint64_t)
+                               << " expected bytes " << bytes;
+                }
+                return false;
+            }
+            cur_page_cache_index = 0;
+        }
+
+        uint64_t page_info = page_cache[cur_page_cache_index++];
+        if (!PAGE_PRESENT(page_info) && !PAGE_SWAPPED(page_info)) continue;
+
+        if (PAGE_SWAPPED(page_info)) {
             vma.usage.swap += pagesz;
-            swap_offsets_.emplace_back(PAGE_SWAP_OFFSET(p));
+            swap_offsets_.emplace_back(PAGE_SWAP_OFFSET(page_info));
             continue;
         }
 
-        uint64_t page_frame = PAGE_PFN(p);
-        if (!pinfo.PageFlags(page_frame, &pg_flags[i])) {
+        uint64_t page_frame = PAGE_PFN(page_info);
+        uint64_t cur_page_flags;
+        if (!pinfo.PageFlags(page_frame, &cur_page_flags)) {
             LOG(ERROR) << "Failed to get page flags for " << page_frame << " in process " << pid_;
             swap_offsets_.clear();
             return false;
         }
 
         // skip unwanted pages from the count
-        if ((pg_flags[i] & pgflags_mask_) != pgflags_) continue;
+        if ((cur_page_flags & pgflags_mask_) != pgflags_) continue;
 
-        if (!pinfo.PageMapCount(page_frame, &pg_counts[i])) {
+        uint64_t cur_page_counts;
+        if (!pinfo.PageMapCount(page_frame, &cur_page_counts)) {
             LOG(ERROR) << "Failed to get page count for " << page_frame << " in process " << pid_;
             swap_offsets_.clear();
             return false;
         }
 
         // Page was unmapped between the presence check at the beginning of the loop and here.
-        if (pg_counts[i] == 0) {
-            pg_frames[i] = 0;
-            pg_flags[i] = 0;
+        if (cur_page_counts == 0) {
             continue;
         }
 
-        bool is_dirty = !!(pg_flags[i] & (1 << KPF_DIRTY));
-        bool is_private = (pg_counts[i] == 1);
+        bool is_dirty = !!(cur_page_flags & (1 << KPF_DIRTY));
+        bool is_private = (cur_page_counts == 1);
         // Working set
         if (get_wss) {
             bool is_referenced = use_pageidle ? (pinfo.IsPageIdle(page_frame) == 1)
-                                              : !!(pg_flags[i] & (1 << KPF_REFERENCED));
+                                              : !!(cur_page_flags & (1 << KPF_REFERENCED));
             if (!is_referenced) {
                 continue;
             }
@@ -351,7 +373,7 @@
 
         vma.usage.rss += pagesz;
         vma.usage.uss += is_private ? pagesz : 0;
-        vma.usage.pss += pagesz / pg_counts[i];
+        vma.usage.pss += pagesz / cur_page_counts;
         if (is_private) {
             vma.usage.private_dirty += is_dirty ? pagesz : 0;
             vma.usage.private_clean += is_dirty ? 0 : pagesz;
diff --git a/libmeminfo/tools/procrank.cpp b/libmeminfo/tools/procrank.cpp
index cb3757d..1e44ff9 100644
--- a/libmeminfo/tools/procrank.cpp
+++ b/libmeminfo/tools/procrank.cpp
@@ -348,7 +348,7 @@
     auto rss_sort = [](ProcessRecord& a, ProcessRecord& b) {
         MemUsage stats_a = show_wss ? a.Wss() : a.Usage();
         MemUsage stats_b = show_wss ? b.Wss() : b.Usage();
-        return reverse_sort ? stats_a.rss < stats_b.pss : stats_a.pss > stats_b.pss;
+        return reverse_sort ? stats_a.rss < stats_b.rss : stats_a.rss > stats_b.rss;
     };
 
     auto vss_sort = [](ProcessRecord& a, ProcessRecord& b) {
diff --git a/libprocessgroup/cgroup_map.cpp b/libprocessgroup/cgroup_map.cpp
index 9797d76..20ae2be 100644
--- a/libprocessgroup/cgroup_map.cpp
+++ b/libprocessgroup/cgroup_map.cpp
@@ -67,11 +67,14 @@
     return controller_ != nullptr;
 }
 
-bool CgroupController::IsUsable() const {
+bool CgroupController::IsUsable() {
     if (!HasValue()) return false;
 
-    static bool enabled = (access(GetProcsFilePath("", 0, 0).c_str(), F_OK) == 0);
-    return enabled;
+    if (state_ == UNKNOWN) {
+        state_ = access(GetProcsFilePath("", 0, 0).c_str(), F_OK) == 0 ? USABLE : MISSING;
+    }
+
+    return state_ == USABLE;
 }
 
 std::string CgroupController::GetTasksFilePath(const std::string& rel_path) const {
diff --git a/libprocessgroup/cgroup_map.h b/libprocessgroup/cgroup_map.h
index 9350412..427d71b 100644
--- a/libprocessgroup/cgroup_map.h
+++ b/libprocessgroup/cgroup_map.h
@@ -31,20 +31,28 @@
 class CgroupController {
   public:
     // Does not own controller
-    explicit CgroupController(const ACgroupController* controller) : controller_(controller) {}
+    explicit CgroupController(const ACgroupController* controller)
+        : controller_(controller), state_(UNKNOWN) {}
 
     uint32_t version() const;
     const char* name() const;
     const char* path() const;
 
     bool HasValue() const;
-    bool IsUsable() const;
+    bool IsUsable();
 
     std::string GetTasksFilePath(const std::string& path) const;
     std::string GetProcsFilePath(const std::string& path, uid_t uid, pid_t pid) const;
     bool GetTaskGroup(int tid, std::string* group) const;
   private:
+    enum ControllerState {
+        UNKNOWN = 0,
+        USABLE = 1,
+        MISSING = 2,
+    };
+
     const ACgroupController* controller_ = nullptr;
+    ControllerState state_;
 };
 
 class CgroupMap {