1 /*
  2  * Copyright (c) 2023, Red Hat, Inc. All rights reserved.
  3  * Copyright (c) 2023, 2024, Oracle and/or its affiliates. All rights reserved.
  4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  5  *
  6  * This code is free software; you can redistribute it and/or modify it
  7  * under the terms of the GNU General Public License version 2 only, as
  8  * published by the Free Software Foundation.
  9  *
 10  * This code is distributed in the hope that it will be useful, but WITHOUT
 11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 13  * version 2 for more details (a copy is included in the LICENSE file that
 14  * accompanied this code).
 15  *
 16  * You should have received a copy of the GNU General Public License version
 17  * 2 along with this work; if not, write to the Free Software Foundation,
 18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 19  *
 20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 21  * or visit www.oracle.com if you need additional information or have any
 22  * questions.
 23  *
 24  */
 25 
 26 #include "precompiled.hpp"
 27 #include "asm/assembler.hpp"
 28 #include "logging/log.hpp"
 29 #include "oops/compressedKlass.hpp"
 30 #include "memory/metaspace.hpp"
 31 #include "runtime/os.hpp"
 32 #include "utilities/globalDefinitions.hpp"
 33 
 34 // Helper function; reserve at an address that is compatible with EOR
 35 static char* reserve_at_eor_compatible_address(size_t size, bool aslr) {
 36   char* result = nullptr;
 37 
 38   log_debug(metaspace, map)("Trying to reserve at an EOR-compatible address");
 39 
 40   // We need immediates that are 32-bit aligned, since they should not intersect nKlass
 41   // bits. They should not be larger than the addressable space either, but we still
 42   // lack a good abstraction for that (see JDK-8320584), therefore we assume and hard-code
 43   // 2^48 as a reasonable higher ceiling.
 44   static const uint16_t immediates[] = {
 45       0x0001, 0x0002, 0x0003, 0x0004, 0x0006, 0x0007, 0x0008, 0x000c, 0x000e,
 46       0x000f, 0x0010, 0x0018, 0x001c, 0x001e, 0x001f, 0x0020, 0x0030, 0x0038,
 47       0x003c, 0x003e, 0x003f, 0x0040, 0x0060, 0x0070, 0x0078, 0x007c, 0x007e,
 48       0x007f, 0x0080, 0x00c0, 0x00e0, 0x00f0, 0x00f8, 0x00fc, 0x00fe, 0x00ff,
 49       0x0100, 0x0180, 0x01c0, 0x01e0, 0x01f0, 0x01f8, 0x01fc, 0x01fe, 0x01ff,
 50       0x0200, 0x0300, 0x0380, 0x03c0, 0x03e0, 0x03f0, 0x03f8, 0x03fc, 0x03fe,
 51       0x03ff, 0x0400, 0x0600, 0x0700, 0x0780, 0x07c0, 0x07e0, 0x07f0, 0x07f8,
 52       0x07fc, 0x07fe, 0x07ff, 0x0800, 0x0c00, 0x0e00, 0x0f00, 0x0f80, 0x0fc0,
 53       0x0fe0, 0x0ff0, 0x0ff8, 0x0ffc, 0x0ffe, 0x0fff, 0x1000, 0x1800, 0x1c00,
 54       0x1e00, 0x1f00, 0x1f80, 0x1fc0, 0x1fe0, 0x1ff0, 0x1ff8, 0x1ffc, 0x1ffe,
 55       0x1fff, 0x2000, 0x3000, 0x3800, 0x3c00, 0x3e00, 0x3f00, 0x3f80, 0x3fc0,
 56       0x3fe0, 0x3ff0, 0x3ff8, 0x3ffc, 0x3ffe, 0x3fff, 0x4000, 0x6000, 0x7000,
 57       0x7800, 0x7c00, 0x7e00, 0x7f00, 0x7f80, 0x7fc0, 0x7fe0, 0x7ff0, 0x7ff8,
 58       0x7ffc, 0x7ffe, 0x7fff
 59   };
 60   static constexpr unsigned num_immediates = sizeof(immediates) / sizeof(immediates[0]);
 61   const unsigned start_index = aslr ? os::next_random((int)os::javaTimeNanos()) : 0;
 62   constexpr int max_tries = 64;
 63   for (int ntry = 0; result == nullptr && ntry < max_tries; ntry ++) {
 64     // As in os::attempt_reserve_memory_between, we alternate between higher and lower
 65     // addresses; this maximizes the chance of early success if part of the address space
 66     // is not accessible (e.g. 39-bit address space).
 67     const unsigned alt_index = (ntry & 1) ? 0 : num_immediates / 2;
 68     const unsigned index = (start_index + ntry + alt_index) % num_immediates;
 69     const uint64_t immediate = ((uint64_t)immediates[index]) << 32;
 70     assert(immediate > 0 && Assembler::operand_valid_for_logical_immediate(/*is32*/false, immediate),
 71            "Invalid immediate %d " UINT64_FORMAT, index, immediate);
 72     result = os::attempt_reserve_memory_at((char*)immediate, size, false);
 73     if (result == nullptr) {
 74       log_trace(metaspace, map)("Failed to attach at " UINT64_FORMAT_X, immediate);
 75     }
 76   }
 77   if (result == nullptr) {
 78     log_debug(metaspace, map)("Failed to reserve at any EOR-compatible address");
 79   }
 80   return result;
 81 }
 82 char* CompressedKlassPointers::reserve_address_space_for_compressed_classes(size_t size, bool aslr, bool optimize_for_zero_base) {
 83 
 84   char* result = nullptr;
 85 
 86   // Optimize for base=0 shift=0
 87   if (optimize_for_zero_base) {
 88     result = reserve_address_space_for_unscaled_encoding(size, aslr);
 89   }
 90 
 91   // If this fails, we don't bother aiming for zero-based encoding (base=0 shift>0), since it has no
 92   // advantages over EOR or movk mode.
 93 
 94   // EOR-compatible reservation
 95   if (result == nullptr) {
 96     result = reserve_at_eor_compatible_address(size, aslr);
 97   }
 98 
 99   // Movk-compatible reservation via probing.
100   if (result == nullptr) {
101     result = reserve_address_space_for_16bit_move(size, aslr);
102   }
103 
104   // Movk-compatible reservation via overallocation.
105   // If that failed, attempt to allocate at any 4G-aligned address. Let the system decide where. For ASLR,
106   // we now rely on the system.
107   // Compared with the probing done above, this has two disadvantages:
108   // - on a kernel with 52-bit address space we may get an address that has bits set between [48, 52).
109   //   In that case, we may need two movk moves (not yet implemented).
110   // - this technique leads to temporary over-reservation of address space; it will spike the vsize of
111   //   the process. Therefore it may fail if a vsize limit is in place (e.g. ulimit -v).
112   if (result == nullptr) {
113     constexpr size_t alignment = nth_bit(32);
114     log_debug(metaspace, map)("Trying to reserve at a 32-bit-aligned address");
115     result = os::reserve_memory_aligned(size, alignment, false);
116   }
117 
118   return result;
119 }
120 
121 bool CompressedKlassPointers::pd_initialize(address addr, size_t len) {
122 
123   if (tiny_classpointer_mode()) {
124     // In tiny-classpointer mode, we do what all other platforms do.
125     return false;
126   }
127 
128   // Aarch64 uses an own initialization logic that avoids zero-base shifted mode
129   // (_base=0 _shift>0), instead preferring non-zero-based mode with shift=0
130   constexpr uintptr_t unscaled_max = nth_bit(32);
131   assert(len <= unscaled_max, "Klass range larger than 32 bits?");
132 
133   _shift = 0;
134 
135   address const end = addr + len;
136   _base = (end <= (address)unscaled_max) ? nullptr : addr;
137 
138   _range = end - _base;
139 
140 #ifdef ASSERT
141   _klass_range_start = addr;
142   _klass_range_end = addr + len;
143   calc_lowest_highest_narrow_klass_id();
144   sanity_check_after_initialization();
145 #endif
146 
147   return true;
148 }