genode/repos/base-foc/patches/0022-FOC-amd64-split-_syscall_entry-into-code-and-data.patch

From f29031cdbe8cebf6c39d02a72dd50c736cec3a69 Mon Sep 17 00:00:00 2001

From: =?UTF-8?q?Jakub=20Jerm=C3=A1=C5=99?= <jakub.jermar@kernkonzept.com>

Subject: [PATCH] amd64: Split _syscall_entry into code and data

This change makes the entire syscall entry table smaller, sparing 32
bytes per a statically configured CPU. Note that some padding is still
used to achieve cacheline alignment of the per-CPU entries of the data
part of the table.

Furthemore, by moving the two data members into a data section and
allowing to modify the handler offset from JDB only, the amd64 kernel is
easier to make completely W^X in the future. The handler entry offset
remains a part of the jump instruction opcode (there is no data
indirection) and JDB is enhanced to be able to patch it when needed.

Change-Id: I22b91f9fd2b108d99e3ceea6611a15ab3db26bb6

Edit for Genode: UX changes not included because of more dependencies
---
 src/jdb/ia32/jdb_trace_set-ia32-ux.cpp |    4 +-
 src/kern/ia32/64/cpu-64.cpp            |   31 +++++-------------
 src/kern/ia32/64/entry.S               |   29 +++++++++++++++++
 src/kern/ia32/64/syscall_entry.cpp     |   54 +++++---------------------------
 4 files changed, 48 insertions(+), 70 deletions(-)

diff --git a/src/jdb/ia32/jdb_trace_set-ia32-ux.cpp b/src/jdb/ia32/jdb_trace_set-ia32-ux.cpp
index ec23322..0a2f309 100644
--- a/src/jdb/ia32/jdb_trace_set-ia32-ux.cpp
+++ b/src/jdb/ia32/jdb_trace_set-ia32-ux.cpp
@@ -39,7 +39,7 @@ Jdb_set_trace::set_ipc_vector()
 
   Idt::set_entry(0x30, (Address) int30_entry, true);
   Jdb::on_each_cpu([fast_entry](Cpu_number cpu){
-    Cpu::cpus.cpu(cpu).set_fast_entry(fast_entry);
+    //Cpu::cpus.cpu(cpu).set_fast_entry(fast_entry);
   });
 
   if (Jdb_ipc_trace::_trace)
@@ -98,7 +98,7 @@ struct Jdb_ipc_log_pm : Pm_object
     else
       fast_entry  = entry_sys_fast_ipc_c;
 
-    Cpu::cpus.cpu(cpu).set_fast_entry(fast_entry);
+    //Cpu::cpus.cpu(cpu).set_fast_entry(fast_entry);
   }
 
   void pm_on_suspend(Cpu_number) override {}
diff --git a/src/kern/ia32/64/cpu-64.cpp b/src/kern/ia32/64/cpu-64.cpp
index 974c677..d18ad16 100644
--- a/src/kern/ia32/64/cpu-64.cpp
+++ b/src/kern/ia32/64/cpu-64.cpp
@@ -4,7 +4,8 @@ INTERFACE [amd64 && !kernel_isolation]:
 
 EXTENSION class Cpu
 {
-  static Per_cpu_array<Syscall_entry> _syscall_entry;
+  static Per_cpu_array<Syscall_entry_data>
+    _syscall_entry_data asm("syscall_entry_data");
 };
 
 
@@ -13,14 +14,7 @@ IMPLEMENTATION[amd64 && !kernel_isolation]:
 #include "mem_layout.h"
 #include "tss.h"
 
-Per_cpu_array<Syscall_entry> Cpu::_syscall_entry;
-
-PUBLIC
-void
-Cpu::set_fast_entry(void (*func)())
-{
-  _syscall_entry[id()].set_entry(func);
-}
+Per_cpu_array<Syscall_entry_data> Cpu::_syscall_entry_data;
 
 IMPLEMENT inline NEEDS["tss.h"]
 Address volatile &
@@ -31,11 +25,13 @@ PUBLIC inline
 void
 Cpu::setup_sysenter()
 {
+  extern Per_cpu_array<Syscall_entry_text> syscall_entry_text;
+
   wrmsr(0, GDT_CODE_KERNEL | ((GDT_CODE_USER32 | 3) << 16), MSR_STAR);
-  wrmsr((Unsigned64)&_syscall_entry[id()], MSR_LSTAR);
-  wrmsr((Unsigned64)&_syscall_entry[id()], MSR_CSTAR);
+  wrmsr((Unsigned64)&syscall_entry_text[id()], MSR_LSTAR);
+  wrmsr((Unsigned64)&syscall_entry_text[id()], MSR_CSTAR);
   wrmsr(~0U, MSR_SFMASK);
-  _syscall_entry[id()].set_rsp((Address)&kernel_sp());
+  _syscall_entry_data[id()].set_rsp((Address)&kernel_sp());
 }
 
 IMPLEMENTATION[amd64 && kernel_isolation]:
@@ -43,16 +39,6 @@ IMPLEMENTATION[amd64 && kernel_isolation]:
 #include "mem_layout.h"
 #include "tss.h"
 
-PUBLIC
-void
-Cpu::set_fast_entry(void (*func)())
-{
-  extern char const syscall_entry_code[];
-  extern char const syscall_entry_reloc[];
-  auto ofs = syscall_entry_reloc - syscall_entry_code + 3; // 3 byte movebas
-  *reinterpret_cast<Signed32 *>(Mem_layout::Mem_layout::Kentry_cpu_page + ofs + 0xa0) = (Signed32)(Signed64)func;
-}
-
 PUBLIC inline
 void
 Cpu::setup_sysenter() const
@@ -78,7 +64,6 @@ Cpu::init_sysenter()
 {
   setup_sysenter();
   wrmsr(rdmsr(MSR_EFER) | 1, MSR_EFER);
-  set_fast_entry(entry_sys_fast_ipc_c);
 }
 
 
diff --git a/src/kern/ia32/64/entry.S b/src/kern/ia32/64/entry.S
index 1cb8137..ed5a04c 100644
--- a/src/kern/ia32/64/entry.S
+++ b/src/kern/ia32/64/entry.S
@@ -372,7 +372,36 @@ entry_\name:
 	jmp	all_syscalls
 .endm
 
+#ifndef CONFIG_KERNEL_ISOLATION
+#ifdef CONFIG_MP
+MAX_NUM_CPUS = CONFIG_MP_MAX_CPUS
+#else
+MAX_NUM_CPUS = 1
+#endif
+
+#define SYSCALL_ENTRY_DATA_SIZE 64
+#define SYSCALL_ENTRY_TEXT_SIZE (0f - 0b)
+#define SYSCALL_ENTRY_OFFSET ((0b - syscall_entry_text) / SYSCALL_ENTRY_TEXT_SIZE)
+#define SYSCALL_ENTRY_DATA (syscall_entry_data + SYSCALL_ENTRY_OFFSET * SYSCALL_ENTRY_DATA_SIZE)
+#define KERN_SP (SYSCALL_ENTRY_DATA + 0)
+#define USER_SP (SYSCALL_ENTRY_DATA + 8)
 	.section ".entry.text.syscalls", "ax", @progbits
+	.global syscall_entry_text
+	.align 64
+syscall_entry_text:
+.rept MAX_NUM_CPUS
+0:
+	mov %rsp, USER_SP(%rip)
+	mov KERN_SP(%rip), %rsp
+	mov (%rsp), %rsp
+	pushq $GDT_DATA_USER | 3
+	pushq USER_SP(%rip)
+	jmp entry_sys_fast_ipc_c
+	.align 32
+0:
+.endr
+#endif /* !CONFIG_KERNEL_ISOLATION */
+
 	.p2align 4
 	.type	all_syscalls,@function
 all_syscalls:
diff --git a/src/kern/ia32/64/syscall_entry.cpp b/src/kern/ia32/64/syscall_entry.cpp
index b03d1ad..3dd7db3 100644
--- a/src/kern/ia32/64/syscall_entry.cpp
+++ b/src/kern/ia32/64/syscall_entry.cpp
@@ -2,58 +2,22 @@ INTERFACE [amd64]:
 
 #include "types.h"
 
-class Syscall_entry
+class Syscall_entry_data
 {
-private:
-  template<int INSN_LEN>
-  struct Mem_insn
-  {
-    Unsigned32 _insn:INSN_LEN * 8;
-    Unsigned32 _offset;
-    Mem_insn(Unsigned32 insn, void *mem)
-    : _insn(insn),
-      _offset((Address)mem - (Address)(&_offset + 1))
-    {}
-  } __attribute__((packed));
-
-  Mem_insn<3> _mov_rsp_user_sp;
-  Mem_insn<3> _mov_kern_sp_rsp;
-  Unsigned32 _mov_rsp_rsp;
-  Unsigned8 _push_ss, _ss_value;
-  Mem_insn<2> _push_user_rsp;
-  Unsigned8 _jmp;
-  Signed32 _entry_offset;
-  Unsigned8 _pading[33]; // pad to the next 64 byte boundary
-  Unsigned64 _kern_sp;
-  Unsigned64 _user_sp;
-} __attribute__((packed, aligned(64)));
+  Unsigned64 _kern_sp = 0;
+  Unsigned64 _user_sp = 0;
+} __attribute__((packed, aligned(64))); // Enforce cacheline alignment
 
+struct Syscall_entry_text
+{
+  char _res[32]; // Keep this in sync with code in syscall_entry_text!
+} __attribute__((packed, aligned(32)));
 
 IMPLEMENTATION [amd64]:
 
-#include "config_gdt.h"
-
-PUBLIC inline NEEDS["config_gdt.h"]
-Syscall_entry::Syscall_entry()
-: /* mov %rsp, _user_sp(%rip) */ _mov_rsp_user_sp(0x258948, &_user_sp),
-  /* mov _kern_sp(%rip), %rsp */ _mov_kern_sp_rsp(0x258b48, &_kern_sp),
-  /* mov (%rsp), %rsp */         _mov_rsp_rsp(0x24248b48),
-  /* pushq GDT_DATA_USER | 3 */  _push_ss(0x6a), _ss_value(GDT_DATA_USER | 3),
-  /* pushq _user_sp(%rip) */     _push_user_rsp(0x35ff, &_user_sp),
-  /* jmp *_entry_offset */       _jmp(0xe9)
-{}
-
-PUBLIC inline
-void
-Syscall_entry::set_entry(void (*func)(void))
-{
-  _entry_offset = (Address)func
-                  - ((Address)&_entry_offset + sizeof(_entry_offset));
-}
-
 PUBLIC inline
 void
-Syscall_entry::set_rsp(Address rsp)
+Syscall_entry_data::set_rsp(Address rsp)
 {
   _kern_sp = rsp;
 }