Skip to contentSkip to conversation

bpf, riscv: use BPF prog pack allocator in BPF JIT

PATCH Superseded Puranjay Mohan 5 participants 21 messages Aug 24, 2023 – Aug 28, 2023

Subsystems (4)

v2 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

v2 cover bpf, riscv: use BPF prog pack allocator in BPF JIT

Changes in v1 -> v2:
1. Implement a new function patch_text_set_nosync() to be used in bpf_arch_text_invalidate(). The implementation in v1 called patch_text_nosync() in a loop and it was bad as it would call flush_icache_range() for every word making it really slow. This was found by running the test_tag selftest which would take forever to complete.

Here is some data to prove the V2 fixes the problem:

Without this series:

root@rv-selftester:~/src/kselftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m47.562s
user    0m24.145s
sys     6m37.064s

With this series applied:

root@rv-selftester:~/src/selftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m29.472s
user    0m25.865s
sys     6m18.401s

BPF programs currently consume a page each on RISCV. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.

Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the x86_64 BPF JIT.

I enabled this allocator on the ARM64 BPF JIT[2]. It is being reviewed now.

This patch series enables the BPF prog pack allocator for the RISCV BPF JIT. This series needs a patch[3] from the ARM64 series to work.

======================================================
Performance Analysis of prog pack allocator on RISCV64
======================================================

Test setup:
===========

Host machine: Debian GNU/Linux 11 (bullseye)
Qemu Version: QEMU emulator version 8.0.3 (Debian 1:8.0.3+dfsg-1) u-boot-qemu Version: 2023.07+dfsg-1
opensbi Version: 1.3-1

To test the performance of the BPF prog pack allocator on RV, a stresser tool[4] linked below was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.

The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered. The script is passed a command which would be run in the above environment.

The script was run with following perf command:

./run.sh "perf stat -a \
        -e iTLB-load-misses \
        -e dTLB-load-misses  \
        -e dTLB-store-misses \
        -e instructions \
        --timeout 60000"

The output of the above command is discussed below before and after enabling the BPF prog pack allocator.

The tests were run on qemu-system-riscv64 with 8 cpus, 16G memory. The rootfs was created using Bjorn's riscv-cross-builder[5] docker container linked below.

Results
=======

Before enabling prog pack allocator:
------------------------------------

Performance counter stats for 'system wide':

           4939048      iTLB-load-misses
           5468689      dTLB-load-misses
            465234      dTLB-store-misses
     1441082097998      instructions
      60.045791200 seconds time elapsed

After enabling prog pack allocator:
-----------------------------------

Performance counter stats for 'system wide':

           3430035      iTLB-load-misses
           5008745      dTLB-load-misses
            409944      dTLB-store-misses
     1441535637988      instructions
      60.046296600 seconds time elapsed

Improvements in metrics
=======================

It was expected that the iTLB-load-misses would decrease as now a single huge page is used to keep all the BPF programs compared to a single page for each program earlier.

--------------------------------------------
The improvement in iTLB-load-misses: -30.5 %
--------------------------------------------

I repeated this expriment more than 100 times in different setups and the improvement was always greater than 30%.

This patch series is boot tested on the Starfive VisionFive 2 board[6]. The performance analysis was not done on the board because it doesn't expose iTLB-load-misses, etc. The stresser program was run on the board to test the loading and unloading of BPF programs

[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/
[4] https://github.com/puranjaymohan/BPF-Allocator-Bench
[5] https://github.com/bjoto/riscv-cross-builder
[6] https://www.starfivetech.com/en/site/boards

Puranjay Mohan (3):
riscv: extend patch_text_nosync() for multiple pages
riscv: implement a memset like function for text bpf, riscv: use prog pack allocator in the BPF JIT

 arch/riscv/include/asm/patch.h  |   1 +
 arch/riscv/kernel/patch.c       | 113 ++++++++++++++++++++++++++++++--
 arch/riscv/net/bpf_jit.h        |   3 +
 arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
 arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
 5 files changed, 255 insertions(+), 31 deletions(-)
Signature · 1 line
2.39.2
#1 lore raw

PATCH v2 1/3 riscv: extend patch_text_nosync() for multiple pages

The patch_insn_write() function currently doesn't work for multiple pages of instructions, therefore patch_text_nosync() will fail with a page fault if called with lengths spanning multiple pages.

This commit extends the patch_insn_write() function to support multiple pages by copying at max 2 pages at a time in a loop. This implementation is similar to text_poke_copy() function of x86.

---
 arch/riscv/kernel/patch.c | 39 ++++++++++++++++++++++++++++++++++-----
 1 file changed, 34 insertions(+), 5 deletions(-)
arch/riscv/kernel/patch.c +34 -5
index 575e71d6c8ae..465b2eebbc37 100644
@@ -53,12 +53,18 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
 	void *waddr = addr;
 	bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
 	int ret;
 
+	/*
+	 * Only two pages can be mapped at a time for writing.
+	 */
+	if (len > 2 * PAGE_SIZE)
+		return -EINVAL;
+
 	/*
 	 * Before reaching here, it was expected to lock the text_mutex
 	 * already, so we don't need to give another lock here and could
@@ -74,7 +80,7 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 		lockdep_assert_held(&text_mutex);
 
 	if (across_pages)
-		patch_map(addr + len, FIX_TEXT_POKE1);
+		patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
 
 	waddr = patch_map(addr, FIX_TEXT_POKE0);
 
@@ -87,15 +93,38 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 
 	return ret;
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #else
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
 	return copy_to_kernel_nofault(addr, insn, len);
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #endif /* CONFIG_MMU */
 
+static int patch_insn_write(void *addr, const void *insn, size_t len)
+{
+	size_t patched = 0;
+	size_t size;
+	int ret = 0;
+
+	/*
+	 * Copy the instructions to the destination address, two pages at a time
+	 * because __patch_insn_write() can only handle len <= 2 * PAGE_SIZE.
+	 */
+	while (patched < len && !ret) {
+		size = min_t(size_t,
+			     PAGE_SIZE * 2 - offset_in_page(addr + patched),
+			     len - patched);
+		ret = __patch_insn_write(addr + patched, insn + patched, size);
+
+		patched += size;
+	}
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_insn_write);
+
 int patch_text_nosync(void *addr, const void *insns, size_t len)
 {
 	u32 *tp = addr;
Signature · 1 line
2.39.2
reviewer
#2 lore raw

Re: [PATCH bpf-next v2 1/3] riscv: extend patch_text_nosync() for multiple pages

Quoted 25 lines · Puranjay Mohan

The patch_insn_write() function currently doesn't work for multiple pages of instructions, therefore patch_text_nosync() will fail with a page fault if called with lengths spanning multiple pages.

This commit extends the patch_insn_write() function to support multiple pages by copying at max 2 pages at a time in a loop. This implementation is similar to text_poke_copy() function of x86.

---
 arch/riscv/kernel/patch.c | 39 ++++++++++++++++++++++++++++++++++-----
 1 file changed, 34 insertions(+), 5 deletions(-)
arch/riscv/kernel/patch.c +1 -1
index 575e71d6c8ae..465b2eebbc37 100644
@@ -53,12 +53,18 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)

arch/riscv/kernel/patch.c static void patch_unmap(int fixmap) · line 66

Quoted 10 lines
arch/riscv/kernel/patch.c +5 -0
 {
        void *waddr = addr;
        bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
        int ret;
 
+       /*
+        * Only two pages can be mapped at a time for writing.
+        */
+       if (len > 2 * PAGE_SIZE)
+               return -EINVAL;

This check cannot guarantee __patch_insn_write touch at most two pages. Maybe use

if (len + offset_in_page(addr) > 2 * PAGE_SIZE)
return -EINVAL;
?

Thanks,
Song

Quoted 57 lines
arch/riscv/kernel/patch.c +27 -5
        /*
         * Before reaching here, it was expected to lock the text_mutex
         * already, so we don't need to give another lock here and could
@@ -74,7 +80,7 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
                lockdep_assert_held(&text_mutex);
 
        if (across_pages)
-               patch_map(addr + len, FIX_TEXT_POKE1);
+               patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
 
        waddr = patch_map(addr, FIX_TEXT_POKE0);
 
@@ -87,15 +93,38 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 
        return ret;
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #else
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
        return copy_to_kernel_nofault(addr, insn, len);
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #endif /* CONFIG_MMU */
 
+static int patch_insn_write(void *addr, const void *insn, size_t len)
+{
+       size_t patched = 0;
+       size_t size;
+       int ret = 0;
+
+       /*
+        * Copy the instructions to the destination address, two pages at a time
+        * because __patch_insn_write() can only handle len <= 2 * PAGE_SIZE.
+        */
+       while (patched < len && !ret) {
+               size = min_t(size_t,
+                            PAGE_SIZE * 2 - offset_in_page(addr + patched),
+                            len - patched);
+               ret = __patch_insn_write(addr + patched, insn + patched, size);
+
+               patched += size;
+       }
+
+       return ret;
+}
+NOKPROBE_SYMBOL(patch_insn_write);
+
 int patch_text_nosync(void *addr, const void *insns, size_t len)
 {
        u32 *tp = addr;
--

2.39.2

#3 lore raw

Re: [PATCH bpf-next v2 1/3] riscv: extend patch_text_nosync() for multiple pages

Hi Song,

Quoted 32 lines · Song Liu
Quoted 30 lines · Puranjay Mohan

The patch_insn_write() function currently doesn't work for multiple pages of instructions, therefore patch_text_nosync() will fail with a page fault if called with lengths spanning multiple pages.

This commit extends the patch_insn_write() function to support multiple pages by copying at max 2 pages at a time in a loop. This implementation is similar to text_poke_copy() function of x86.

---
 arch/riscv/kernel/patch.c | 39 ++++++++++++++++++++++++++++++++++-----
 1 file changed, 34 insertions(+), 5 deletions(-)
arch/riscv/kernel/patch.c +1 -1
index 575e71d6c8ae..465b2eebbc37 100644
@@ -53,12 +53,18 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
        void *waddr = addr;
        bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
        int ret;
 
Quoted 7 lines

arch/riscv/kernel/patch.c static void patch_unmap(int fixmap) · line 66

Quoted 5 lines
arch/riscv/kernel/patch.c +5 -0
+       /*
+        * Only two pages can be mapped at a time for writing.
+        */
+       if (len > 2 * PAGE_SIZE)
+               return -EINVAL;

This check cannot guarantee __patch_insn_write touch at most two pages.

Yes, I just realised this can span 3 pages if len = 2 * PAGE_SIZE and offset_in_page(addr) > 0.

Quoted 5 lines

Maybe use

if (len + offset_in_page(addr) > 2 * PAGE_SIZE)
return -EINVAL;
?

Will fix it in the next version.

Quoted 53 lines

Thanks,
Song

Quoted 49 lines
arch/riscv/kernel/patch.c +25 -4
        /*
         * Before reaching here, it was expected to lock the text_mutex
         * already, so we don't need to give another lock here and could
@@ -74,7 +80,7 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
                lockdep_assert_held(&text_mutex);
 
        if (across_pages)
-               patch_map(addr + len, FIX_TEXT_POKE1);
+               patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
 
        waddr = patch_map(addr, FIX_TEXT_POKE0);
 
@@ -87,15 +93,38 @@ static int patch_insn_write(void *addr, const void *insn, size_t len)
 
        return ret;
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #else
-static int patch_insn_write(void *addr, const void *insn, size_t len)
+static int __patch_insn_write(void *addr, const void *insn, size_t len)
 {
        return copy_to_kernel_nofault(addr, insn, len);
 }
-NOKPROBE_SYMBOL(patch_insn_write);
+NOKPROBE_SYMBOL(__patch_insn_write);
 #endif /* CONFIG_MMU */
 
+static int patch_insn_write(void *addr, const void *insn, size_t len)
+{
+       size_t patched = 0;
+       size_t size;
+       int ret = 0;
+
+       /*
+        * Copy the instructions to the destination address, two pages at a time
+        * because __patch_insn_write() can only handle len <= 2 * PAGE_SIZE.
+        */
+       while (patched < len && !ret) {
+               size = min_t(size_t,
+                            PAGE_SIZE * 2 - offset_in_page(addr + patched),
+                            len - patched);
+               ret = __patch_insn_write(addr + patched, insn + patched, size);
+
+               patched += size;
+       }
+
+       return ret;
+}
Quoted 8 lines
Quoted 8 lines
arch/riscv/kernel/patch.c +2 -1
+NOKPROBE_SYMBOL(patch_insn_write);
+
 int patch_text_nosync(void *addr, const void *insns, size_t len)
 {
        u32 *tp = addr;
--

2.39.2

Thanks,
Puranjay

#4 lore raw

PATCH v2 2/3 riscv: implement a memset like function for text

In reply to Puranjay Mohan

The BPF JIT needs to write invalid instructions to RX regions of memory to invalidate removed BPF programs. This needs a function like memset() that can work with RX memory.

Implement patch_text_set_nosync() which is similar to text_poke_set() of x86.

---
 arch/riscv/include/asm/patch.h |  1 +
 arch/riscv/kernel/patch.c      | 74 ++++++++++++++++++++++++++++++++++
 2 files changed, 75 insertions(+)
arch/riscv/include/asm/patch.h +1 -0
index 63c98833d510..aa5c1830ea43 100644
@@ -7,6 +7,7 @@
 #define _ASM_RISCV_PATCH_H
 
 int patch_text_nosync(void *addr, const void *insns, size_t len);
+int patch_text_set_nosync(void *addr, const int c, size_t len);
 int patch_text(void *addr, u32 *insns, int ninsns);
 
 extern int riscv_patch_in_stop_machine;
arch/riscv/kernel/patch.c +74 -0

98 lines not on this page · raw message

Signature · 1 line
2.39.2
reviewer
#5 lore raw

Re: [PATCH bpf-next v2 2/3] riscv: implement a memset like function for text

Quoted 43 lines · Puranjay Mohan

The BPF JIT needs to write invalid instructions to RX regions of memory to invalidate removed BPF programs. This needs a function like memset() that can work with RX memory.

Implement patch_text_set_nosync() which is similar to text_poke_set() of x86.

---
 arch/riscv/include/asm/patch.h |  1 +
 arch/riscv/kernel/patch.c      | 74 ++++++++++++++++++++++++++++++++++
 2 files changed, 75 insertions(+)
arch/riscv/include/asm/patch.h +1 -0
index 63c98833d510..aa5c1830ea43 100644
@@ -7,6 +7,7 @@
 #define _ASM_RISCV_PATCH_H
 
 int patch_text_nosync(void *addr, const void *insns, size_t len);
+int patch_text_set_nosync(void *addr, const int c, size_t len);
 int patch_text(void *addr, u32 *insns, int ninsns);
 
 extern int riscv_patch_in_stop_machine;
arch/riscv/kernel/patch.c +2 -0
index 465b2eebbc37..24d49999ac1a 100644
@@ -13,6 +13,7 @@
 #include <asm/fixmap.h>
 #include <asm/ftrace.h>
 #include <asm/patch.h>
+#include <asm/string.h>
 
 struct patch_insn {
        void *addr;
@@ -53,6 +54,34 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
+static int __patch_insn_set(void *addr, const int c, size_t len)

arch/riscv/kernel/patch.c static void patch_unmap(int fixmap) · line 67

Quoted 10 lines
arch/riscv/kernel/patch.c +10 -0
+{
+       void *waddr = addr;
+       bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
+       int ret;
+
+       /*
+        * Only two pages can be mapped at a time for writing.
+        */
+       if (len > 2 * PAGE_SIZE)
+               return -EINVAL;

Same for this one.

[...]

#6 lore raw

Re: [PATCH bpf-next v2 2/3] riscv: implement a memset like function for text

In reply to Puranjay Mohan
kernel test robot noticed the following build warnings:

Hi Puranjay,

kernel test robot noticed the following build warnings:

[auto build test WARNING on bpf-next/master]

url:    https://github.com/intel-lab-lkp/linux/commits/Puranjay-Mohan/riscv-extend-patch_text_nosync-for-multiple-pages/20230824-213410
base:   https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git master
patch link:    https://lore.kernel.org/r/20230824133135.1176709-3-puranjay12%40gmail.com
patch subject: [PATCH bpf-next v2 2/3] riscv: implement a memset like function for text
config: riscv-allnoconfig (https://download.01.org/0day-ci/archive/20230825/[email protected]/config)
compiler: riscv64-linux-gcc (GCC) 13.2.0
reproduce: (https://download.01.org/0day-ci/archive/20230825/[email protected]/reproduce)

If you fix the issue in a separate patch/commit (i.e. not just a new version of the same patch/commit), kindly add following tags
| Reported-by: kernel test robot <[email protected]>
| Closes: https://lore.kernel.org/oe-kbuild-all/[email protected]/

All warnings (new ones prefixed by >>):

arch/riscv/kernel/patch.c: In function '__patch_insn_set':

>> arch/riscv/kernel/patch.c:61:13: warning: unused variable 'ret' [-Wunused-variable]
      61 |         int ret;
         |             ^~~

vim +/ret +61 arch/riscv/kernel/patch.c

    56	
    57	static int __patch_insn_set(void *addr, const int c, size_t len)
    58	{
    59		void *waddr = addr;
    60		bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
  > 61		int ret;
    62	
    63		/*
    64		 * Only two pages can be mapped at a time for writing.
    65		 */
    66		if (len > 2 * PAGE_SIZE)
    67			return -EINVAL;
    68	
    69		if (across_pages)
    70			patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
    71	
    72		waddr = patch_map(addr, FIX_TEXT_POKE0);
    73	
    74		memset(waddr, c, len);
    75	
    76		patch_unmap(FIX_TEXT_POKE0);
    77	
    78		if (across_pages)
    79			patch_unmap(FIX_TEXT_POKE1);
    80	
    81		return 0;
    82	}
    83	NOKPROBE_SYMBOL(__patch_insn_set);
    84	
Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests/wiki
maintainer
#7 lore raw

Re: [PATCH bpf-next v2 2/3] riscv: implement a memset like function for text

In reply to Puranjay Mohan
Quoted 43 lines · Puranjay Mohan

The BPF JIT needs to write invalid instructions to RX regions of memory to invalidate removed BPF programs. This needs a function like memset() that can work with RX memory.

Implement patch_text_set_nosync() which is similar to text_poke_set() of x86.

---
  arch/riscv/include/asm/patch.h |  1 +
  arch/riscv/kernel/patch.c      | 74 ++++++++++++++++++++++++++++++++++
  2 files changed, 75 insertions(+)
arch/riscv/include/asm/patch.h +1 -0
index 63c98833d510..aa5c1830ea43 100644
@@ -7,6 +7,7 @@
  #define _ASM_RISCV_PATCH_H
 
  int patch_text_nosync(void *addr, const void *insns, size_t len);
+int patch_text_set_nosync(void *addr, const int c, size_t len);
  int patch_text(void *addr, u32 *insns, int ninsns);
 
  extern int riscv_patch_in_stop_machine;
arch/riscv/kernel/patch.c +3 -0
index 465b2eebbc37..24d49999ac1a 100644
@@ -13,6 +13,7 @@
  #include <asm/fixmap.h>
  #include <asm/ftrace.h>
  #include <asm/patch.h>
+#include <asm/string.h>
 
  struct patch_insn {
  	void *addr;
@@ -53,6 +54,34 @@ static void patch_unmap(int fixmap)
  }
  NOKPROBE_SYMBOL(patch_unmap);
 
+static int __patch_insn_set(void *addr, const int c, size_t len)
+{

arch/riscv/kernel/patch.c static void patch_unmap(int fixmap) · line 68

Quoted 10 lines
arch/riscv/kernel/patch.c +10 -0
+	void *waddr = addr;
+	bool across_pages = (((uintptr_t) addr & ~PAGE_MASK) + len) > PAGE_SIZE;
+	int ret;
+
+	/*
+	 * Only two pages can be mapped at a time for writing.
+	 */
+	if (len > 2 * PAGE_SIZE)
+		return -EINVAL;
+

As a generic part, it better to add text_mutex lock assert.

Quoted 78 lines
arch/riscv/kernel/patch.c +61 -0
+	if (across_pages)
+		patch_map(addr + PAGE_SIZE, FIX_TEXT_POKE1);
+
+	waddr = patch_map(addr, FIX_TEXT_POKE0);
+
+	memset(waddr, c, len);
+
+	patch_unmap(FIX_TEXT_POKE0);
+
+	if (across_pages)
+		patch_unmap(FIX_TEXT_POKE1);
+
+	return 0;
+}
+NOKPROBE_SYMBOL(__patch_insn_set);
+
  static int __patch_insn_write(void *addr, const void *insn, size_t len)
  {
  	void *waddr = addr;
@@ -95,6 +124,14 @@ static int __patch_insn_write(void *addr, const void *insn, size_t len)
  }
  NOKPROBE_SYMBOL(__patch_insn_write);
  #else
+static int __patch_insn_set (void *addr, const int c, size_t len)
+{
+	memset(addr, c, len);
+
+	return 0;
+}
+NOKPROBE_SYMBOL(__patch_insn_set);
+
  static int __patch_insn_write(void *addr, const void *insn, size_t len)
  {
  	return copy_to_kernel_nofault(addr, insn, len);
@@ -102,6 +139,43 @@ static int __patch_insn_write(void *addr, const void *insn, size_t len)
  NOKPROBE_SYMBOL(__patch_insn_write);
  #endif /* CONFIG_MMU */
 
+static int patch_insn_set(void *addr, const int c, size_t len)
+{
+	size_t patched = 0;
+	size_t size;
+	int ret = 0;
+
+	/*
+	 * __patch_insn_set() can only work on 2 pages at a time so call it in a
+	 * loop with len <= 2 * PAGE_SIZE.
+	 */
+	while (patched < len && !ret) {
+		size = min_t(size_t,
+			     PAGE_SIZE * 2 - offset_in_page(addr + patched),
+			     len - patched);
+		ret = __patch_insn_set(addr + patched, c, size);
+
+		patched += size;
+	}
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_insn_set);
+
+int patch_text_set_nosync(void *addr, const int c, size_t len)
+{
+	u32 *tp = addr;
+	int ret;
+
+	ret = patch_insn_set(tp, c, len);
+
+	if (!ret)
+		flush_icache_range((uintptr_t) tp, (uintptr_t) tp + len);
+
+	return ret;
+}
+NOKPROBE_SYMBOL(patch_text_set_nosync);
+
  static int patch_insn_write(void *addr, const void *insn, size_t len)
  {
  	size_t patched = 0;
maintainer
#8 lore raw

Re: [PATCH bpf-next v2 2/3] riscv: implement a memset like function for text

In reply to Puranjay Mohan
Quoted 6 lines · Puranjay Mohan

The BPF JIT needs to write invalid instructions to RX regions of memory to invalidate removed BPF programs. This needs a function like memset() that can work with RX memory.

Implement patch_text_set_nosync() which is similar to text_poke_set() of x86.

Some additional nits, in addition to the other comments (Song, kernel test bot, Lehui).

Quoted 31 lines
---
 arch/riscv/include/asm/patch.h |  1 +
 arch/riscv/kernel/patch.c      | 74 ++++++++++++++++++++++++++++++++++
 2 files changed, 75 insertions(+)
arch/riscv/include/asm/patch.h +1 -0
index 63c98833d510..aa5c1830ea43 100644
@@ -7,6 +7,7 @@
 #define _ASM_RISCV_PATCH_H
 
 int patch_text_nosync(void *addr, const void *insns, size_t len);
+int patch_text_set_nosync(void *addr, const int c, size_t len);
 int patch_text(void *addr, u32 *insns, int ninsns);
 
 extern int riscv_patch_in_stop_machine;
arch/riscv/kernel/patch.c +1 -0
index 465b2eebbc37..24d49999ac1a 100644
@@ -13,6 +13,7 @@
 #include <asm/fixmap.h>
 #include <asm/ftrace.h>
 #include <asm/patch.h>
+#include <asm/string.h>
 
 struct patch_insn {
 	void *addr;

arch/riscv/kernel/patch.c static void patch_unmap(int fixmap) · line 57

Quoted 4 lines
arch/riscv/kernel/patch.c +1 -0
@@ -53,6 +54,34 @@ static void patch_unmap(int fixmap)
 }
 NOKPROBE_SYMBOL(patch_unmap);
 
+static int __patch_insn_set(void *addr, const int c, size_t len)

Drop the "const" from "const int c" everywhere in this patch, and let's just use u8 instead of int. We don't need to carry the old memset() legacy argumentts! We're more modern than that! ;-)

Björn

#9 lore raw

PATCH v2 3/3 bpf, riscv: use prog pack allocator in the BPF JIT

In reply to Puranjay Mohan

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
 arch/riscv/net/bpf_jit.h        |   3 +
 arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
 arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
 3 files changed, 146 insertions(+), 26 deletions(-)
arch/riscv/net/bpf_jit.h +3 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
 struct rv_jit_context {
 	struct bpf_prog *prog;
 	u16 *insns;		/* RV insns */
+	u16 *ro_insns;
 	int ninsns;
 	int prologue_len;
 	int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
 
 struct rv_jit_data {
 	struct bpf_binary_header *header;
+	struct bpf_binary_header *ro_header;
 	u8 *image;
+	u8 *ro_image;
 	struct rv_jit_context ctx;
 };
 
arch/riscv/net/bpf_jit_comp64.c +46 -10

104 lines not on this page · raw message

arch/riscv/net/bpf_jit_core.c +97 -16

165 lines not on this page · raw message

Signature · 1 line
2.39.2
reviewer Reviewed-by on v2 3/3
#10 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Quoted 11 lines · Puranjay Mohan

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

LGTM.

[...]

maintainer
#11 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

In reply to Puranjay Mohan

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

Quoted 56 lines · Puranjay Mohan

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
  arch/riscv/net/bpf_jit.h        |   3 +
  arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
  arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
  3 files changed, 146 insertions(+), 26 deletions(-)
arch/riscv/net/bpf_jit.h +3 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
  struct rv_jit_context {
  	struct bpf_prog *prog;
  	u16 *insns;		/* RV insns */
+	u16 *ro_insns;
  	int ninsns;
  	int prologue_len;
  	int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
 
  struct rv_jit_data {
  	struct bpf_binary_header *header;
+	struct bpf_binary_header *ro_header;
  	u8 *image;
+	u8 *ro_image;
  	struct rv_jit_context ctx;
  };
 
arch/riscv/net/bpf_jit_comp64.c +5 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
  /* Emit fixed-length instructions for address */
  static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct rv_jit_context *ctx)
  {
-	u64 ip = (u64)(ctx->insns + ctx->ninsns);
+	/*
+	 * Use the ro_insns(RX) to calculate the offset as the BPF program will
+	 * finally run from this memory region.
+	 */
+	u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
  	s64 off = addr - ip;
  	s64 upper = (off + (1 << 11)) >> 12;
  	s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, struct rv_jit_context *ctx) · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr, struct rv_jit_context *ctx)
  	u64 ip;
 
  	if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 10 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-		ip = (u64)(long)(ctx->insns + ctx->ninsns);
+		/*
+		 * Use the ro_insns(RX) to calculate the offset as the BPF
+		 * program will finally run from this memory region.
+		 */
+		ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
  		off = addr - ip;
  	}
 

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct bpf_insn *insn, · line 592

Quoted 8 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct bpf_insn *insn,
  {
  	struct exception_table_entry *ex;
  	unsigned long pc;
-	off_t offset;
+	off_t ins_offset;
+	off_t fixup_offset;
 
  	if (!ctx->insns || !ctx->prog->aux->extable || BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 210 lines
arch/riscv/net/bpf_jit_comp64.c +34 -7
  		return 0;
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct bpf_insn *insn,
  		return -EINVAL;
 
  	ex = &ctx->prog->aux->extable[ctx->nexentries];
-	pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+	pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
 
-	offset = pc - (long)&ex->insn;
-	if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+	/*
+	 * This is the relative offset of the instruction that may fault from
+	 * the exception table itself. This will be written to the exception
+	 * table and if this instruction faults, the destination register will
+	 * be set to '0' and the execution will jump to the next instruction.
+	 */
+	ins_offset = pc - (long)&ex->insn;
+	if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
  		return -ERANGE;
-	ex->insn = offset;
 
  	/*
  	 * Since the extable follows the program, the fixup offset is always
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct bpf_insn *insn,
  	 * bits. We don't need to worry about buildtime or runtime sort
  	 * modifying the upper bits because the table is already sorted, and
  	 * isn't part of the main exception table.
+	 *
+	 * The fixup_offset is set to the next instruction from the instruction
+	 * that may fault. The execution will jump to this after handling the
+	 * fault.
  	 */
-	offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+	fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
  		return -ERANGE;
 
-	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+	/*
+	 * The offsets above have been calculated using the RO buffer but we
+	 * need to use the R/W buffer for writes.
+	 * switch ex to rw buffer for writing.
+	 */
+	ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+	ex->insn = ins_offset;
+
+	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
  		FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
  	ex->type = EX_TYPE_BPF;
 
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
 
  	ctx.ninsns = 0;
  	ctx.insns = NULL;
+	ctx.ro_insns = NULL;
  	ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx);
  	if (ret < 0)
  		return ret;
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
  		return -EFBIG;
 
  	ctx.ninsns = 0;
+	/*
+	 * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to write the
+	 * JITed instructions and later copies it to a RX region (ctx.ro_insns).
+	 * It also uses ctx.ro_insns to calculate offsets for jumps etc. As the
+	 * trampoline image uses the same memory area for writing and execution,
+	 * both ctx.insns and ctx.ro_insns can be set to image.
+	 */
  	ctx.insns = image;
+	ctx.ro_insns = image;
  	ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx);
  	if (ret < 0)
  		return ret;
arch/riscv/net/bpf_jit_core.c +56 -16
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
 
  #include <linux/bpf.h>
  #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
  #include "bpf_jit.h"
 
  /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
  				sizeof(struct exception_table_entry);
  			prog_size = sizeof(*ctx->insns) * ctx->ninsns;
 
-			jit_data->header =
-				bpf_jit_binary_alloc(prog_size + extable_size,
-						     &jit_data->image,
-						     sizeof(u32),
-						     bpf_fill_ill_insns);
-			if (!jit_data->header) {
+			jit_data->ro_header =
+				bpf_jit_binary_pack_alloc(prog_size +
+							  extable_size,
+							  &jit_data->ro_image,
+							  sizeof(u32),
+							  &jit_data->header,
+							  &jit_data->image,
+							  bpf_fill_ill_insns);
+			if (!jit_data->ro_header) {
  				prog = orig_prog;
  				goto out_offset;
  			}
 
+			/*
+			 * Use the image(RW) for writing the JITed instructions. But also save
+			 * the ro_image(RX) for calculating the offsets in the image. The RW
+			 * image will be later copied to the RX image from where the program
+			 * will run. The bpf_jit_binary_pack_finalize() will do this copy in the
+			 * final step.
+			 */
+			ctx->ro_insns = (u16 *)jit_data->ro_image;
  			ctx->insns = (u16 *)jit_data->image;
  			/*
  			 * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 
  	if (i == NR_JIT_ITERATIONS) {
  		pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-		if (jit_data->header)
-			bpf_jit_binary_free(jit_data->header);
  		prog = orig_prog;
-		goto out_offset;
+		goto out_free_hdr;
  	}
 
  	if (extable_size)
-		prog->aux->extable = (void *)ctx->insns + prog_size;
+		prog->aux->extable = (void *)ctx->ro_insns + prog_size;
 
  skip_init_ctx:
  	pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
 
  	bpf_jit_build_prologue(ctx);
  	if (build_body(ctx, extra_pass, NULL)) {
-		bpf_jit_binary_free(jit_data->header);
  		prog = orig_prog;
-		goto out_offset;
+		goto out_free_hdr;
  	}
  	bpf_jit_build_epilogue(ctx);
 
  	if (bpf_jit_enable > 1)
  		bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
 
-	prog->bpf_func = (void *)ctx->insns;
+	prog->bpf_func = (void *)ctx->ro_insns;
  	prog->jited = 1;
  	prog->jited_len = prog_size;
 
-	bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
  	if (!prog->is_func || extra_pass) {
-		bpf_jit_binary_lock_ro(jit_data->header);
+		if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+							 jit_data->ro_header,
+							 jit_data->header))) {
+			/* ro_header has been freed */
+			jit_data->ro_header = NULL;
+			prog = orig_prog;
+			goto out_offset;
+		}
+		/*
+		 * The instructions have now been copied to the ROX region from
+		 * where they will execute.
+		 * Write any modified data cache blocks out to memory and
+		 * invalidate the corresponding blocks in the instruction cache.
+		 */
+		bpf_flush_icache(jit_data->ro_header,
+				 ctx->ro_insns + ctx->ninsns);
  		for (i = 0; i < prog->len; i++)
  			ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
  		bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
  		bpf_jit_prog_release_other(prog, prog == orig_prog ?
  					   tmp : orig_prog);
  	return prog;
+
+out_free_hdr:
+	if (jit_data->header) {
+		bpf_arch_text_copy(&jit_data->ro_header->size,
+				   &jit_data->header->size,
+				   sizeof(jit_data->header->size));
+		bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+	}
+	goto out_offset;
  }
 
  u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
  {
  	return vfree(addr);
  }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+	int ret;
+
+	mutex_lock(&text_mutex);
+	ret = patch_text_nosync(dst, src, len);
+	mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+	if (ret)
+		return ERR_PTR(-EINVAL);
+
+	return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+	int ret = 0;

no need to initialize it

Quoted 31 lines
arch/riscv/net/bpf_jit_core.c +31 -0
+
+	mutex_lock(&text_mutex);
+	ret = patch_text_set_nosync(dst, 0, len);
+	mutex_unlock(&text_mutex);
+
+	return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+	if (prog->jited) {
+		struct rv_jit_data *jit_data = prog->aux->jit_data;
+		struct bpf_binary_header *hdr;
+
+		/*
+		 * If we fail the final pass of JIT (from jit_subprogs),
+		 * the program may not be finalized yet. Call finalize here
+		 * before freeing it.
+		 */
+		if (jit_data) {
+			bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+						     jit_data->header);
+			kfree(jit_data);
+		}
+		hdr = bpf_jit_binary_pack_hdr(prog);
+		bpf_jit_binary_pack_free(hdr, NULL);
+		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+	}
+
+	bpf_prog_unlock_free(prog);
+}
maintainer
#12 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
  arch/riscv/net/bpf_jit.h        |   3 + arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++--- arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----

  3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 9 lines
Quoted 9 lines
arch/riscv/net/bpf_jit.h +0 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
   struct rv_jit_context {
       struct bpf_prog *prog;

      u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

And any other parts, it looks great.😄

Quoted 314 lines
Quoted 30 lines

      int ninsns;
      int prologue_len;
      int epilogue_offset;

arch/riscv/net/bpf_jit.h +2 -0
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
   struct rv_jit_data {
       struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
       u8 *image;
+    u8 *ro_image;
       struct rv_jit_context ctx;
   };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
   /* Emit fixed-length instructions for address */
   static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct rv_jit_context *ctx)
   {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF 

program will
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
      s64 off = addr - ip;
      s64 upper = (off + (1 << 11)) >> 12;
      s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr, 
 struct rv_jit_context *ctx)
       u64 ip;
       if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 9 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
           off = addr - ip;
       }

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct · line 592

Quoted 8 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct 
 bpf_insn *insn,
   {
       struct exception_table_entry *ex;
       unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
       if (!ctx->insns || !ctx->prog->aux->extable || BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 212 lines
arch/riscv/net/bpf_jit_comp64.c +3 -3
           return 0;
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct 
 bpf_insn *insn,
           return -EINVAL;
       ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault 

from
+     * the exception table itself. This will be written to the exception +     * table and if this instruction faults, the destination register will
+     * be set to '0' and the execution will jump to the next instruction.
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN)) return -ERANGE;
-    ex->insn = offset;
      /*
       * Since the extable follows the program, the fixup offset is always

arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct 
 bpf_insn *insn,
        * bits. We don't need to worry about buildtime or runtime sort
        * modifying the upper bits because the table is already sorted, and
        * isn't part of the main exception table.
+     *
+     * The fixup_offset is set to the next instruction from the 

instruction
+     * that may fault. The execution will jump to this after handling the
+     * fault.
       */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16)); -    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16)); +    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset)) return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) | +    /*
+     * The offsets above have been calculated using the RO buffer but we +     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);

arch/riscv/net/bpf_jit_comp64.c +4 -0
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |

          FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg); ex->type = EX_TYPE_BPF;

arch/riscv/net/bpf_jit_comp64.c +3 -0
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct 
 bpf_tramp_image *im, void *image,
       ctx.ninsns = 0;
       ctx.insns = NULL;
+    ctx.ro_insns = NULL;
       ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx);
       if (ret < 0)
           return ret;
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct 
 bpf_tramp_image *im, void *image,
           return -EFBIG;
       ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to 

write the
+     * JITed instructions and later copies it to a RX region (ctx.ro_insns).
+     * It also uses ctx.ro_insns to calculate offsets for jumps etc. As the
+     * trampoline image uses the same memory area for writing and execution,
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
      ctx.insns = image;
+    ctx.ro_insns = image;
      ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr, flags, &ctx); if (ret < 0)
          return ret;

arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
   #include <linux/bpf.h>
   #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
   #include "bpf_jit.h"
   /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct 
 bpf_prog *prog)
                   sizeof(struct exception_table_entry);
               prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                   prog = orig_prog;
                   goto out_offset;
               }
+            /*
+             * Use the image(RW) for writing the JITed instructions. 

But also save
+             * the ro_image(RX) for calculating the offsets in the image. The RW
+             * image will be later copied to the RX image from where the program
+             * will run. The bpf_jit_binary_pack_finalize() will do this copy in the
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image; ctx->insns = (u16 *)jit_data->image;
              /*
               * Now, when the image is allocated, the image can

arch/riscv/net/bpf_jit_core.c +2 -4
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct 
 bpf_prog *prog)
       if (i == NR_JIT_ITERATIONS) {
           pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;

  skip_init_ctx:
      pass++;

arch/riscv/net/bpf_jit_core.c +18 -6
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct 
 bpf_prog *prog)
       bpf_jit_build_prologue(ctx);
       if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       bpf_jit_build_epilogue(ctx);
       if (bpf_jit_enable > 1)
           bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
       prog->jited = 1;
       prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
       if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
           for (i = 0; i < prog->len; i++)
               ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
           bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct 
 bpf_prog *prog)

          bpf_jit_prog_release_other(prog, prog == orig_prog ? tmp : orig_prog);
      return prog;

arch/riscv/net/bpf_jit_core.c +9 -0
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;

  }
  u64 bpf_jit_alloc_exec_limit(void)

arch/riscv/net/bpf_jit_core.c +8 -0
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
   {
       return vfree(addr);
   }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 31 lines
arch/riscv/net/bpf_jit_core.c +31 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}
#13 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Hi Pu,

Quoted 28 lines · Pu Lehui
Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
  arch/riscv/net/bpf_jit.h        |   3 +
  arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
  arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
  3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 16 lines
Quoted 9 lines

arch/riscv/net/bpf_jit.h static inline bool is_creg(u8 reg) · line 71

Quoted 9 lines
arch/riscv/net/bpf_jit.h +1 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
  struct rv_jit_context {
      struct bpf_prog *prog;
      u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

I had thought about this and decided to create a new _ro image/header and not _rw image/header. Here is my reasoning:
If we let the existing insns, header be considered the read_only version from where the
program will run, and create new rw_insn and rw_header for doing the jit process it would require a lot more changes to the framework.
functions like build_body(), bpf_jit_build_prologue(), etc. work on ctx->insns and
now all these references would have to be changed to ctx->rw_insns.

Howsoever we implement this, there is no way to do it without changing the current framework.
The crux of the problem is that we need to use the r/w area for writing and the r/x area for calculating
offsets.

If you think this can be done in a more efficient way then I would love to implement that, but all other
solutions that I tried made the code very difficult to follow.

Quoted 40 lines

And any other parts, it looks great.😄

Quoted 37 lines
Quoted 32 lines
arch/riscv/net/bpf_jit.h +2 -0
      int ninsns;
      int prologue_len;
      int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
  struct rv_jit_data {
      struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
      u8 *image;
+    u8 *ro_image;
      struct rv_jit_context ctx;
  };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
  /* Emit fixed-length instructions for address */
  static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct
 rv_jit_context *ctx)
  {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF

program will

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
      s64 off = addr - ip;
      s64 upper = (off + (1 << 11)) >> 12;
      s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr,
 struct rv_jit_context *ctx)
      u64 ip;
      if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Can you elaborate this more. I am missing something here. The sync happens at the end by calling bpf_jit_binary_pack_finalize().

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
          off = addr - ip;
      }
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
  {
      struct exception_table_entry *ex;
Quoted 8 lines
Quoted 8 lines

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct bpf_insn *insn,

Quoted 6 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
      unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
      if (!ctx->insns || !ctx->prog->aux->extable ||
 BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

ctx->ro_insns is not initialised until we call bpf_jit_binary_pack_finalize()?

Quoted 256 lines
Quoted 256 lines
Quoted 218 lines
          return 0;
arch/riscv/net/bpf_jit_comp64.c +3 -3
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
          return -EINVAL;
      ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault

from

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     * the exception table itself. This will be written to the exception
+     * table and if this instruction faults, the destination register

will
+ * be set to '0' and the execution will jump to the next
instruction.

arch/riscv/net/bpf_jit_comp64.c +3 -1
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
          return -ERANGE;
-    ex->insn = offset;
      /*
       * Since the extable follows the program, the fixup offset is

always

arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
       * bits. We don't need to worry about buildtime or runtime sort
       * modifying the upper bits because the table is already sorted,

and
* isn't part of the main exception table.

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     *
+     * The fixup_offset is set to the next instruction from the

instruction
+ * that may fault. The execution will jump to this after handling
the

arch/riscv/net/bpf_jit_comp64.c +14 -3
+     * fault.
       */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
          return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+    /*
+     * The offsets above have been calculated using the RO buffer but we
+     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
          FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
      ex->type = EX_TYPE_BPF;
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
      ctx.ninsns = 0;
      ctx.insns = NULL;
+    ctx.ro_insns = NULL;
      ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
 flags, &ctx);
      if (ret < 0)
          return ret;
arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
          return -EFBIG;
      ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to

write the
+ * JITed instructions and later copies it to a RX region
(ctx.ro_insns).
+ * It also uses ctx.ro_insns to calculate offsets for jumps etc.
As the
+ * trampoline image uses the same memory area for writing and
execution,

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
      ctx.insns = image;
+    ctx.ro_insns = image;
      ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
flags, &ctx);
      if (ret < 0)
          return ret;
arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
  #include <linux/bpf.h>
  #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
  #include "bpf_jit.h"
  /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
                  sizeof(struct exception_table_entry);
              prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                  prog = orig_prog;
                  goto out_offset;
              }
+            /*
+             * Use the image(RW) for writing the JITed instructions.

But also save
+ * the ro_image(RX) for calculating the offsets in the
image. The RW
+ * image will be later copied to the RX image from where
the program
+ * will run. The bpf_jit_binary_pack_finalize() will do
this copy in the

arch/riscv/net/bpf_jit_core.c +40 -10
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image;
              ctx->insns = (u16 *)jit_data->image;
              /*
               * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
      if (i == NR_JIT_ITERATIONS) {
          pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
          prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
      }
      if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;
  skip_init_ctx:
      pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
      bpf_jit_build_prologue(ctx);
      if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
          prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
      }
      bpf_jit_build_epilogue(ctx);
      if (bpf_jit_enable > 1)
          bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
      prog->jited = 1;
      prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
      if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
          for (i = 0; i < prog->len; i++)
              ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
          bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
          bpf_jit_prog_release_other(prog, prog == orig_prog ?
                         tmp : orig_prog);
      return prog;
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;
  }
  u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
  {
      return vfree(addr);
  }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 25 lines
arch/riscv/net/bpf_jit_core.c +25 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
Quoted 8 lines
Quoted 8 lines

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 287

Quoted 6 lines
arch/riscv/net/bpf_jit_core.c +6 -0
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}

Thanks,
Puranjay

_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv

maintainer
#14 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Quoted 48 lines · Puranjay Mohan

Hi Pu,

Quoted 28 lines · Pu Lehui
Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
   arch/riscv/net/bpf_jit.h        |   3 +
   arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
   arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
   3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 16 lines
Quoted 9 lines

arch/riscv/net/bpf_jit.h static inline bool is_creg(u8 reg) · line 71

Quoted 9 lines
arch/riscv/net/bpf_jit.h +1 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
   struct rv_jit_context {
       struct bpf_prog *prog;
       u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

Quoted 8 lines

I had thought about this and decided to create a new _ro image/header and not _rw image/header. Here is my reasoning:
If we let the existing insns, header be considered the read_only version from where the
program will run, and create new rw_insn and rw_header for doing the jit process it would require a lot more changes to the framework.
functions like build_body(), bpf_jit_build_prologue(), etc. work on ctx->insns and

Hmm, the other parts should be fine, but the emit instruction is a problem. All right, let's go ahead.

Quoted 53 lines

now all these references would have to be changed to ctx->rw_insns.

Howsoever we implement this, there is no way to do it without changing the current framework.
The crux of the problem is that we need to use the r/w area for writing and the r/x area for calculating
offsets.

If you think this can be done in a more efficient way then I would love to implement that, but all other
solutions that I tried made the code very difficult to follow.

Quoted 40 lines

And any other parts, it looks great.😄

Quoted 37 lines
Quoted 32 lines
arch/riscv/net/bpf_jit.h +2 -0
       int ninsns;
       int prologue_len;
       int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
   struct rv_jit_data {
       struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
       u8 *image;
+    u8 *ro_image;
       struct rv_jit_context ctx;
   };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
   /* Emit fixed-length instructions for address */
   static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct
 rv_jit_context *ctx)
   {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF

program will

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
       s64 off = addr - ip;
       s64 upper = (off + (1 << 11)) >> 12;
       s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr,
 struct rv_jit_context *ctx)
       u64 ip;
       if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 2 lines

Can you elaborate this more. I am missing something here. The sync happens at the end by calling bpf_jit_binary_pack_finalize().

if (addr && ctx->insns) {
	ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
	off = addr - ip;
}

emit ctx->insns + off

Here we are assuming ctx->insns == ctx->ro_insns, if they not, the offset calculated by ctx->ro_insns will not meaningful for ctx->insns. I was curious why we need to use ro_insns to calculate offset? Is that any problem if we do jit iteration with ctx->insns and the final copy ctx->insns to ro_insns?

Quoted 23 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
           off = addr - ip;
       }
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
   {
       struct exception_table_entry *ex;
Quoted 8 lines
Quoted 8 lines

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct bpf_insn *insn,

Quoted 6 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
       unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
       if (!ctx->insns || !ctx->prog->aux->extable ||
 BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 1 line

ctx->ro_insns is not initialised until we call bpf_jit_binary_pack_finalize()?

if (!ctx->insns || !ctx->prog->aux->extable ||
...
pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];

The uninitialized ctx->ro_insns may lead to illegal address access. Although it will never happen, because we also assume that ctx->insns == ctx->ro_insns.

Quoted 268 lines
Quoted 257 lines
Quoted 257 lines
Quoted 219 lines

           return 0;
arch/riscv/net/bpf_jit_comp64.c +3 -3
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
           return -EINVAL;
       ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault

from

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     * the exception table itself. This will be written to the exception
+     * table and if this instruction faults, the destination register

will
+ * be set to '0' and the execution will jump to the next
instruction.

arch/riscv/net/bpf_jit_comp64.c +3 -1
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
           return -ERANGE;
-    ex->insn = offset;
       /*
        * Since the extable follows the program, the fixup offset is

always

arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
        * bits. We don't need to worry about buildtime or runtime sort
        * modifying the upper bits because the table is already sorted,

and
* isn't part of the main exception table.

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     *
+     * The fixup_offset is set to the next instruction from the

instruction
+ * that may fault. The execution will jump to this after handling
the

arch/riscv/net/bpf_jit_comp64.c +14 -3
+     * fault.
        */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
           return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+    /*
+     * The offsets above have been calculated using the RO buffer but we
+     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
           FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
       ex->type = EX_TYPE_BPF;
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
       ctx.ninsns = 0;
       ctx.insns = NULL;
+    ctx.ro_insns = NULL;
       ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
 flags, &ctx);
       if (ret < 0)
           return ret;
arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
           return -EFBIG;
       ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to

write the
+ * JITed instructions and later copies it to a RX region
(ctx.ro_insns).
+ * It also uses ctx.ro_insns to calculate offsets for jumps etc.
As the
+ * trampoline image uses the same memory area for writing and
execution,

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
       ctx.insns = image;
+    ctx.ro_insns = image;
       ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
flags, &ctx);
       if (ret < 0)
           return ret;
arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
   #include <linux/bpf.h>
   #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
   #include "bpf_jit.h"
   /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
                   sizeof(struct exception_table_entry);
               prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                   prog = orig_prog;
                   goto out_offset;
               }
+            /*
+             * Use the image(RW) for writing the JITed instructions.

But also save
+ * the ro_image(RX) for calculating the offsets in the
image. The RW
+ * image will be later copied to the RX image from where
the program
+ * will run. The bpf_jit_binary_pack_finalize() will do
this copy in the

arch/riscv/net/bpf_jit_core.c +40 -10
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image;
               ctx->insns = (u16 *)jit_data->image;
               /*
                * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
       if (i == NR_JIT_ITERATIONS) {
           pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;
   skip_init_ctx:
       pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
       bpf_jit_build_prologue(ctx);
       if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       bpf_jit_build_epilogue(ctx);
       if (bpf_jit_enable > 1)
           bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
       prog->jited = 1;
       prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
       if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
           for (i = 0; i < prog->len; i++)
               ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
           bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
           bpf_jit_prog_release_other(prog, prog == orig_prog ?
                          tmp : orig_prog);
       return prog;
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;
   }
   u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
   {
       return vfree(addr);
   }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 25 lines
arch/riscv/net/bpf_jit_core.c +25 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
Quoted 8 lines
Quoted 8 lines

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 287

Quoted 6 lines
arch/riscv/net/bpf_jit_core.c +6 -0
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}

Thanks,
Puranjay

#15 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Hi Pu,

Quoted 120 lines · Pu Lehui
Quoted 48 lines · Puranjay Mohan

Hi Pu,

Quoted 28 lines · Pu Lehui
Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
   arch/riscv/net/bpf_jit.h        |   3 +
   arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
   arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
   3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 16 lines
Quoted 9 lines

arch/riscv/net/bpf_jit.h static inline bool is_creg(u8 reg) · line 71

Quoted 9 lines
arch/riscv/net/bpf_jit.h +1 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
   struct rv_jit_context {
       struct bpf_prog *prog;
       u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

Quoted 8 lines

I had thought about this and decided to create a new _ro image/header and not _rw image/header. Here is my reasoning:
If we let the existing insns, header be considered the read_only version from where the
program will run, and create new rw_insn and rw_header for doing the jit process it would require a lot more changes to the framework.
functions like build_body(), bpf_jit_build_prologue(), etc. work on ctx->insns and

Hmm, the other parts should be fine, but the emit instruction is a problem. All right, let's go ahead.

Quoted 53 lines

now all these references would have to be changed to ctx->rw_insns.

Howsoever we implement this, there is no way to do it without changing the current framework.
The crux of the problem is that we need to use the r/w area for writing and the r/x area for calculating
offsets.

If you think this can be done in a more efficient way then I would love to implement that, but all other
solutions that I tried made the code very difficult to follow.

Quoted 40 lines

And any other parts, it looks great.😄

Quoted 37 lines
Quoted 32 lines
arch/riscv/net/bpf_jit.h +2 -0
       int ninsns;
       int prologue_len;
       int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
   struct rv_jit_data {
       struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
       u8 *image;
+    u8 *ro_image;
       struct rv_jit_context ctx;
   };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
   /* Emit fixed-length instructions for address */
   static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct
 rv_jit_context *ctx)
   {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF

program will

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
       s64 off = addr - ip;
       s64 upper = (off + (1 << 11)) >> 12;
       s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr,
 struct rv_jit_context *ctx)
       u64 ip;
       if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 2 lines

Can you elaborate this more. I am missing something here. The sync happens at the end by calling bpf_jit_binary_pack_finalize().

Quoted 8 lines
if (addr && ctx->insns) {
        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
        off = addr - ip;
}

emit ctx->insns + off

Here we are assuming ctx->insns == ctx->ro_insns, if they not, the offset calculated by ctx->ro_insns will not meaningful for ctx->insns.

We are not assuming that ctx->insns == ctx->ro_insns at this point. We are just finding the offset: off = addr(let's say in kernel) - ip(address of the instruction);

Quoted 3 lines

I was curious why we need to use ro_insns to calculate offset? Is that any problem if we do jit iteration with ctx->insns and the final copy ctx->insns to ro_insns?

All the offsets within the image can be calculated using ctx->insns and it will work but if the emit_call() is for an address in the kernel code let's say, then the
offset between this address(in kernel) and the R/W image would be different from the offset between the address(in kernel) and the R/O image.
We need the offset between the R/X Image and the kernel address. Because the CPU will execute the instructions from there.

Quoted 24 lines
Quoted 24 lines
Quoted 23 lines
Quoted 23 lines
Quoted 12 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
           off = addr - ip;
       }

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct · line 593

Quoted 9 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
   {
       struct exception_table_entry *ex;
       unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
       if (!ctx->insns || !ctx->prog->aux->extable ||
 BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 1 line
Quoted 1 line

ctx->ro_insns is not initialised until we call bpf_jit_binary_pack_finalize()?

ctx->ro_insns and ctx->insns are both allocated together by bpf_jit_binary_pack_alloc().
ctx->ro_insns is marked R/X and ctx->insns is marked R/W. We dump all instructions in
ctx->insns and then copy them to ctx->ro_insns with bpf_jit_binary_pack_finalize().

The catch is that instructions that work with offsets like JAL need the offsets from ctx->ro_insns.
as explained above.

Quoted 8 lines

if (!ctx->insns || !ctx->prog->aux->extable ||
...
pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];

The uninitialized ctx->ro_insns may lead to illegal address access. Although it will never happen, because we also assume that ctx->insns == ctx->ro_insns.

Here also we are not assuming ctx->insns == ctx->ro_insns. The ctx->ro_insns is allocated but not initialised yet. So all addresses in range ctx->ro_insns to ctx->ro_insns + size
are valid addresses. Here we are using the addresses only to find the offset and not accessing those
addresses.

Quoted 267 lines
Quoted 267 lines
Quoted 266 lines
Quoted 266 lines
Quoted 220 lines


           return 0;
arch/riscv/net/bpf_jit_comp64.c +3 -3
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
           return -EINVAL;
       ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault

from

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     * the exception table itself. This will be written to the exception
+     * table and if this instruction faults, the destination register

will
+ * be set to '0' and the execution will jump to the next
instruction.

arch/riscv/net/bpf_jit_comp64.c +3 -1
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
           return -ERANGE;
-    ex->insn = offset;
       /*
        * Since the extable follows the program, the fixup offset is

always

arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
        * bits. We don't need to worry about buildtime or runtime sort
        * modifying the upper bits because the table is already sorted,

and
* isn't part of the main exception table.

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     *
+     * The fixup_offset is set to the next instruction from the

instruction
+ * that may fault. The execution will jump to this after handling
the

arch/riscv/net/bpf_jit_comp64.c +14 -3
+     * fault.
        */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
           return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+    /*
+     * The offsets above have been calculated using the RO buffer but we
+     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
           FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
       ex->type = EX_TYPE_BPF;
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
       ctx.ninsns = 0;
       ctx.insns = NULL;
+    ctx.ro_insns = NULL;
       ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
 flags, &ctx);
       if (ret < 0)
           return ret;
arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
           return -EFBIG;
       ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to

write the
+ * JITed instructions and later copies it to a RX region
(ctx.ro_insns).
+ * It also uses ctx.ro_insns to calculate offsets for jumps etc.
As the
+ * trampoline image uses the same memory area for writing and
execution,

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
       ctx.insns = image;
+    ctx.ro_insns = image;
       ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
flags, &ctx);
       if (ret < 0)
           return ret;
arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
   #include <linux/bpf.h>
   #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
   #include "bpf_jit.h"
   /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
                   sizeof(struct exception_table_entry);
               prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                   prog = orig_prog;
                   goto out_offset;
               }
+            /*
+             * Use the image(RW) for writing the JITed instructions.

But also save
+ * the ro_image(RX) for calculating the offsets in the
image. The RW
+ * image will be later copied to the RX image from where
the program
+ * will run. The bpf_jit_binary_pack_finalize() will do
this copy in the

arch/riscv/net/bpf_jit_core.c +40 -10
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image;
               ctx->insns = (u16 *)jit_data->image;
               /*
                * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
       if (i == NR_JIT_ITERATIONS) {
           pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;
   skip_init_ctx:
       pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
       bpf_jit_build_prologue(ctx);
       if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
           prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
       }
       bpf_jit_build_epilogue(ctx);
       if (bpf_jit_enable > 1)
           bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
       prog->jited = 1;
       prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
       if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
           for (i = 0; i < prog->len; i++)
               ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
           bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
           bpf_jit_prog_release_other(prog, prog == orig_prog ?
                          tmp : orig_prog);
       return prog;
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;
   }
   u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
   {
       return vfree(addr);
   }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 31 lines
arch/riscv/net/bpf_jit_core.c +31 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}
Quoted 2 lines
Quoted 2 lines

Thanks,
Puranjay

Thanks,
Puranjay

maintainer
#16 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Quoted 140 lines · Puranjay Mohan

Hi Pu,

Quoted 120 lines · Pu Lehui
Quoted 48 lines · Puranjay Mohan

Hi Pu,

Quoted 28 lines · Pu Lehui
Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
    arch/riscv/net/bpf_jit.h        |   3 +
    arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
    arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
    3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 16 lines
Quoted 9 lines

arch/riscv/net/bpf_jit.h static inline bool is_creg(u8 reg) · line 71

Quoted 9 lines
arch/riscv/net/bpf_jit.h +1 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
    struct rv_jit_context {
        struct bpf_prog *prog;
        u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

Quoted 8 lines

I had thought about this and decided to create a new _ro image/header and not _rw image/header. Here is my reasoning:
If we let the existing insns, header be considered the read_only version from where the
program will run, and create new rw_insn and rw_header for doing the jit process it would require a lot more changes to the framework.
functions like build_body(), bpf_jit_build_prologue(), etc. work on ctx->insns and

Hmm, the other parts should be fine, but the emit instruction is a problem. All right, let's go ahead.

Quoted 53 lines

now all these references would have to be changed to ctx->rw_insns.

Howsoever we implement this, there is no way to do it without changing the current framework.
The crux of the problem is that we need to use the r/w area for writing and the r/x area for calculating
offsets.

If you think this can be done in a more efficient way then I would love to implement that, but all other
solutions that I tried made the code very difficult to follow.

Quoted 40 lines

And any other parts, it looks great.😄

Quoted 37 lines
Quoted 32 lines
arch/riscv/net/bpf_jit.h +2 -0
        int ninsns;
        int prologue_len;
        int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
    struct rv_jit_data {
        struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
        u8 *image;
+    u8 *ro_image;
        struct rv_jit_context ctx;
    };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
    /* Emit fixed-length instructions for address */
    static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct
 rv_jit_context *ctx)
    {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF

program will

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
        s64 off = addr - ip;
        s64 upper = (off + (1 << 11)) >> 12;
        s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr,
 struct rv_jit_context *ctx)
        u64 ip;
        if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 2 lines

Can you elaborate this more. I am missing something here. The sync happens at the end by calling bpf_jit_binary_pack_finalize().

Quoted 8 lines
if (addr && ctx->insns) {
         ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
         off = addr - ip;
}

emit ctx->insns + off

Here we are assuming ctx->insns == ctx->ro_insns, if they not, the offset calculated by ctx->ro_insns will not meaningful for ctx->insns.

We are not assuming that ctx->insns == ctx->ro_insns at this point. We are just finding the offset: off = addr(let's say in kernel) - ip(address of the instruction);

Quoted 3 lines

I was curious why we need to use ro_insns to calculate offset? Is that any problem if we do jit iteration with ctx->insns and the final copy ctx->insns to ro_insns?

Quoted 7 lines

All the offsets within the image can be calculated using ctx->insns and it will work but if the emit_call() is for an address in the kernel code let's say, then the
offset between this address(in kernel) and the R/W image would be different from the offset between the address(in kernel) and the R/O image.
We need the offset between the R/X Image and the kernel address. Because the CPU will execute the instructions from there.

Agree with that, thanks for explaination. Let's talk about my original idea, shall we add check like this to reject ctx->ro_insns == NULL?

if (addr && ctx->insns && ctx->ro_insns) {
...
}

Quoted 34 lines
Quoted 25 lines
Quoted 25 lines
Quoted 24 lines
Quoted 24 lines
Quoted 13 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
            off = addr - ip;
        }

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct · line 593

Quoted 9 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
    {
        struct exception_table_entry *ex;
        unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
        if (!ctx->insns || !ctx->prog->aux->extable ||
 BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 1 line
Quoted 1 line

ctx->ro_insns is not initialised until we call bpf_jit_binary_pack_finalize()?

ctx->ro_insns and ctx->insns are both allocated together by bpf_jit_binary_pack_alloc().
ctx->ro_insns is marked R/X and ctx->insns is marked R/W. We dump all instructions in
ctx->insns and then copy them to ctx->ro_insns with bpf_jit_binary_pack_finalize().

Quoted 8 lines

The catch is that instructions that work with offsets like JAL need the offsets from ctx->ro_insns.
as explained above.

Quoted 4 lines

if (!ctx->insns || !ctx->prog->aux->extable ||
...
pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];

Also here, to add check like this to reject ctx->ro_insns == NULL which may cause null pointer dereference?

if (!ctx->insns || !ctx->ro_insns || !ctx->prog->aux->extable ||

Quoted 285 lines
Quoted 4 lines

The uninitialized ctx->ro_insns may lead to illegal address access. Although it will never happen, because we also assume that ctx->insns == ctx->ro_insns.

Here also we are not assuming ctx->insns == ctx->ro_insns. The ctx->ro_insns is allocated but not initialised yet. So all addresses in range ctx->ro_insns to ctx->ro_insns + size
are valid addresses. Here we are using the addresses only to find the offset and not accessing those
addresses.

Quoted 267 lines
Quoted 267 lines
Quoted 266 lines
Quoted 266 lines
Quoted 220 lines


            return 0;
arch/riscv/net/bpf_jit_comp64.c +3 -3
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
            return -EINVAL;
        ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault

from

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     * the exception table itself. This will be written to the exception
+     * table and if this instruction faults, the destination register

will
+ * be set to '0' and the execution will jump to the next
instruction.

arch/riscv/net/bpf_jit_comp64.c +3 -1
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
            return -ERANGE;
-    ex->insn = offset;
        /*
         * Since the extable follows the program, the fixup offset is

always

arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
         * bits. We don't need to worry about buildtime or runtime sort
         * modifying the upper bits because the table is already sorted,

and
* isn't part of the main exception table.

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     *
+     * The fixup_offset is set to the next instruction from the

instruction
+ * that may fault. The execution will jump to this after handling
the

arch/riscv/net/bpf_jit_comp64.c +14 -3
+     * fault.
         */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
            return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+    /*
+     * The offsets above have been calculated using the RO buffer but we
+     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
            FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
        ex->type = EX_TYPE_BPF;
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
        ctx.ninsns = 0;
        ctx.insns = NULL;
+    ctx.ro_insns = NULL;
        ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
 flags, &ctx);
        if (ret < 0)
            return ret;
arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
            return -EFBIG;
        ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to

write the
+ * JITed instructions and later copies it to a RX region
(ctx.ro_insns).
+ * It also uses ctx.ro_insns to calculate offsets for jumps etc.
As the
+ * trampoline image uses the same memory area for writing and
execution,

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
        ctx.insns = image;
+    ctx.ro_insns = image;
        ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
flags, &ctx);
        if (ret < 0)
            return ret;
arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
    #include <linux/bpf.h>
    #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
    #include "bpf_jit.h"
    /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
                    sizeof(struct exception_table_entry);
                prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                    prog = orig_prog;
                    goto out_offset;
                }
+            /*
+             * Use the image(RW) for writing the JITed instructions.

But also save
+ * the ro_image(RX) for calculating the offsets in the
image. The RW
+ * image will be later copied to the RX image from where
the program
+ * will run. The bpf_jit_binary_pack_finalize() will do
this copy in the

arch/riscv/net/bpf_jit_core.c +40 -10
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image;
                ctx->insns = (u16 *)jit_data->image;
                /*
                 * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
        if (i == NR_JIT_ITERATIONS) {
            pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
            prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
        }
        if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;
    skip_init_ctx:
        pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
        bpf_jit_build_prologue(ctx);
        if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
            prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
        }
        bpf_jit_build_epilogue(ctx);
        if (bpf_jit_enable > 1)
            bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
        prog->jited = 1;
        prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
        if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
            for (i = 0; i < prog->len; i++)
                ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
            bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
            bpf_jit_prog_release_other(prog, prog == orig_prog ?
                           tmp : orig_prog);
        return prog;
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;
    }
    u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
    {
        return vfree(addr);
    }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 31 lines
arch/riscv/net/bpf_jit_core.c +31 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}
Quoted 2 lines
Quoted 2 lines

Thanks,
Puranjay

Thanks,
Puranjay

#17 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

Hi Pu,

Quoted 152 lines · Pu Lehui
Quoted 140 lines · Puranjay Mohan

Hi Pu,

Quoted 120 lines · Pu Lehui
Quoted 48 lines · Puranjay Mohan

Hi Pu,

Quoted 28 lines · Pu Lehui
Quoted 3 lines · Pu Lehui

Hi Puranjay,

Happy to see the RV64 pack allocator implementation.

RV32 also

Quoted 18 lines · Puranjay Mohan
Quoted 16 lines

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

---
    arch/riscv/net/bpf_jit.h        |   3 +
    arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
    arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
    3 files changed, 146 insertions(+), 26 deletions(-)
Quoted 16 lines
Quoted 9 lines

arch/riscv/net/bpf_jit.h static inline bool is_creg(u8 reg) · line 71

Quoted 9 lines
arch/riscv/net/bpf_jit.h +1 -0
index 2717f5490428..ad69319c8ea7 100644
@@ -68,6 +68,7 @@ static inline bool is_creg(u8 reg)
    struct rv_jit_context {
        struct bpf_prog *prog;
        u16 *insns;        /* RV insns */
+    u16 *ro_insns;

In fact, the definition of w/ or w/o ro_ still looks a bit confusing. Maybe it is better for us not to change the current framework, as the current image is the final executed RX image, and the trampoline treats image as the same. Maybe it would be better to add a new RW image, such like rw_iamge, so that we do not break the existing framework and do not have to add too many comments.

Quoted 8 lines

I had thought about this and decided to create a new _ro image/header and not _rw image/header. Here is my reasoning:
If we let the existing insns, header be considered the read_only version from where the
program will run, and create new rw_insn and rw_header for doing the jit process it would require a lot more changes to the framework.
functions like build_body(), bpf_jit_build_prologue(), etc. work on ctx->insns and

Hmm, the other parts should be fine, but the emit instruction is a problem. All right, let's go ahead.

Quoted 53 lines

now all these references would have to be changed to ctx->rw_insns.

Howsoever we implement this, there is no way to do it without changing the current framework.
The crux of the problem is that we need to use the r/w area for writing and the r/x area for calculating
offsets.

If you think this can be done in a more efficient way then I would love to implement that, but all other
solutions that I tried made the code very difficult to follow.

Quoted 40 lines

And any other parts, it looks great.😄

Quoted 37 lines
Quoted 32 lines
arch/riscv/net/bpf_jit.h +2 -0
        int ninsns;
        int prologue_len;
        int epilogue_offset;
@@ -85,7 +86,9 @@ static inline int ninsns_rvoff(int ninsns)
    struct rv_jit_data {
        struct bpf_binary_header *header;
+    struct bpf_binary_header *ro_header;
        u8 *image;
+    u8 *ro_image;
        struct rv_jit_context ctx;
    };
arch/riscv/net/bpf_jit_comp64.c +2 -1
index 0ca4f5c0097c..d77b16338ba2 100644
@@ -144,7 +144,11 @@ static bool in_auipc_jalr_range(s64 val)
    /* Emit fixed-length instructions for address */
    static int emit_addr(u8 rd, u64 addr, bool extra_pass, struct
 rv_jit_context *ctx)
    {
-    u64 ip = (u64)(ctx->insns + ctx->ninsns);
+    /*
+     * Use the ro_insns(RX) to calculate the offset as the BPF

program will

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * finally run from this memory region.
+     */
+    u64 ip = (u64)(ctx->ro_insns + ctx->ninsns);
        s64 off = addr - ip;
        s64 upper = (off + (1 << 11)) >> 12;
        s64 lower = off & 0xfff;

arch/riscv/net/bpf_jit_comp64.c static int emit_call(u64 addr, bool fixed_addr, · line 471

Quoted 3 lines
arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -465,7 +469,11 @@ static int emit_call(u64 addr, bool fixed_addr,
 struct rv_jit_context *ctx)
        u64 ip;
        if (addr && ctx->insns) {

ctx->insns need to sync to ctx->ro_insns

Quoted 2 lines

Can you elaborate this more. I am missing something here. The sync happens at the end by calling bpf_jit_binary_pack_finalize().

Quoted 8 lines
if (addr && ctx->insns) {
         ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
         off = addr - ip;
}

emit ctx->insns + off

Here we are assuming ctx->insns == ctx->ro_insns, if they not, the offset calculated by ctx->ro_insns will not meaningful for ctx->insns.

We are not assuming that ctx->insns == ctx->ro_insns at this point. We are just finding the offset: off = addr(let's say in kernel) - ip(address of the instruction);

Quoted 3 lines

I was curious why we need to use ro_insns to calculate offset? Is that any problem if we do jit iteration with ctx->insns and the final copy ctx->insns to ro_insns?

Quoted 7 lines

All the offsets within the image can be calculated using ctx->insns and it will work but if the emit_call() is for an address in the kernel code let's say, then the
offset between this address(in kernel) and the R/W image would be different from the offset between the address(in kernel) and the R/O image.
We need the offset between the R/X Image and the kernel address. Because the CPU will execute the instructions from there.

Quoted 6 lines

Agree with that, thanks for explaination. Let's talk about my original idea, shall we add check like this to reject ctx->ro_insns == NULL?

if (addr && ctx->insns && ctx->ro_insns) {
...
}

Will add in the next version.

Quoted 40 lines
Quoted 40 lines
Quoted 26 lines
Quoted 26 lines
Quoted 25 lines
Quoted 25 lines
Quoted 14 lines
arch/riscv/net/bpf_jit_comp64.c +5 -1
-        ip = (u64)(long)(ctx->insns + ctx->ninsns);
+        /*
+         * Use the ro_insns(RX) to calculate the offset as the BPF
+         * program will finally run from this memory region.
+         */
+        ip = (u64)(long)(ctx->ro_insns + ctx->ninsns);
            off = addr - ip;
        }

arch/riscv/net/bpf_jit_comp64.c static int add_exception_handler(const struct · line 593

Quoted 9 lines
arch/riscv/net/bpf_jit_comp64.c +2 -1
@@ -578,7 +586,8 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
    {
        struct exception_table_entry *ex;
        unsigned long pc;
-    off_t offset;
+    off_t ins_offset;
+    off_t fixup_offset;
        if (!ctx->insns || !ctx->prog->aux->extable ||
 BPF_MODE(insn->code) != BPF_PROBE_MEM)

ctx->ro_insns need to be checked also.

Quoted 1 line
Quoted 1 line

ctx->ro_insns is not initialised until we call bpf_jit_binary_pack_finalize()?

ctx->ro_insns and ctx->insns are both allocated together by bpf_jit_binary_pack_alloc().
ctx->ro_insns is marked R/X and ctx->insns is marked R/W. We dump all instructions in
ctx->insns and then copy them to ctx->ro_insns with bpf_jit_binary_pack_finalize().

The catch is that instructions that work with offsets like JAL need the offsets from ctx->ro_insns.
as explained above.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

if (!ctx->insns || !ctx->prog->aux->extable ||
...
pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];

Also here, to add check like this to reject ctx->ro_insns == NULL which may cause null pointer dereference?

if (!ctx->insns || !ctx->ro_insns || !ctx->prog->aux->extable ||

Will add in next version.

Quoted 286 lines
Quoted 286 lines
Quoted 5 lines

The uninitialized ctx->ro_insns may lead to illegal address access. Although it will never happen, because we also assume that ctx->insns == ctx->ro_insns.

Here also we are not assuming ctx->insns == ctx->ro_insns. The ctx->ro_insns is allocated but not initialised yet. So all addresses in range ctx->ro_insns to ctx->ro_insns + size
are valid addresses. Here we are using the addresses only to find the offset and not accessing those
addresses.

Quoted 267 lines
Quoted 267 lines
Quoted 266 lines
Quoted 266 lines
Quoted 220 lines


            return 0;
arch/riscv/net/bpf_jit_comp64.c +3 -3
@@ -593,12 +602,17 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
            return -EINVAL;
        ex = &ctx->prog->aux->extable[ctx->nexentries];
-    pc = (unsigned long)&ctx->insns[ctx->ninsns - insn_len];
+    pc = (unsigned long)&ctx->ro_insns[ctx->ninsns - insn_len];
-    offset = pc - (long)&ex->insn;
-    if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
+    /*
+     * This is the relative offset of the instruction that may fault

from

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     * the exception table itself. This will be written to the exception
+     * table and if this instruction faults, the destination register

will
+ * be set to '0' and the execution will jump to the next
instruction.

arch/riscv/net/bpf_jit_comp64.c +3 -1
+     */
+    ins_offset = pc - (long)&ex->insn;
+    if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
            return -ERANGE;
-    ex->insn = offset;
        /*
         * Since the extable follows the program, the fixup offset is

always

arch/riscv/net/bpf_jit_comp64.c +0 -0
@@ -607,12 +621,25 @@ static int add_exception_handler(const struct
 bpf_insn *insn,
         * bits. We don't need to worry about buildtime or runtime sort
         * modifying the upper bits because the table is already sorted,

and
* isn't part of the main exception table.

arch/riscv/net/bpf_jit_comp64.c +2 -0
+     *
+     * The fixup_offset is set to the next instruction from the

instruction
+ * that may fault. The execution will jump to this after handling
the

arch/riscv/net/bpf_jit_comp64.c +14 -3
+     * fault.
         */
-    offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
-    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
+    fixup_offset = (long)&ex->fixup - (pc + insn_len * sizeof(u16));
+    if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
            return -ERANGE;
-    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
+    /*
+     * The offsets above have been calculated using the RO buffer but we
+     * need to use the R/W buffer for writes.
+     * switch ex to rw buffer for writing.
+     */
+    ex = (void *)ctx->insns + ((void *)ex - (void *)ctx->ro_insns);
+
+    ex->insn = ins_offset;
+
+    ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
            FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
        ex->type = EX_TYPE_BPF;
@@ -1006,6 +1033,7 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
        ctx.ninsns = 0;
        ctx.insns = NULL;
+    ctx.ro_insns = NULL;
        ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
 flags, &ctx);
        if (ret < 0)
            return ret;
arch/riscv/net/bpf_jit_comp64.c +2 -0
@@ -1014,7 +1042,15 @@ int arch_prepare_bpf_trampoline(struct
 bpf_tramp_image *im, void *image,
            return -EFBIG;
        ctx.ninsns = 0;
+    /*
+     * The bpf_int_jit_compile() uses a RW buffer (ctx.insns) to

write the
+ * JITed instructions and later copies it to a RX region
(ctx.ro_insns).
+ * It also uses ctx.ro_insns to calculate offsets for jumps etc.
As the
+ * trampoline image uses the same memory area for writing and
execution,

arch/riscv/net/bpf_jit_comp64.c +3 -0
+     * both ctx.insns and ctx.ro_insns can be set to image.
+     */
        ctx.insns = image;
+    ctx.ro_insns = image;
        ret = __arch_prepare_bpf_trampoline(im, m, tlinks, func_addr,
flags, &ctx);
        if (ret < 0)
            return ret;
arch/riscv/net/bpf_jit_core.c +13 -6
index 7a26a3e1c73c..4c8dffc09368 100644
@@ -8,6 +8,8 @@
    #include <linux/bpf.h>
    #include <linux/filter.h>
+#include <linux/memory.h>
+#include <asm/patch.h>
    #include "bpf_jit.h"
    /* Number of iterations to try until offsets converge. */
@@ -117,16 +119,27 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
                    sizeof(struct exception_table_entry);
                prog_size = sizeof(*ctx->insns) * ctx->ninsns;
-            jit_data->header =
-                bpf_jit_binary_alloc(prog_size + extable_size,
-                             &jit_data->image,
-                             sizeof(u32),
-                             bpf_fill_ill_insns);
-            if (!jit_data->header) {
+            jit_data->ro_header =
+                bpf_jit_binary_pack_alloc(prog_size +
+                              extable_size,
+                              &jit_data->ro_image,
+                              sizeof(u32),
+                              &jit_data->header,
+                              &jit_data->image,
+                              bpf_fill_ill_insns);
+            if (!jit_data->ro_header) {
                    prog = orig_prog;
                    goto out_offset;
                }
+            /*
+             * Use the image(RW) for writing the JITed instructions.

But also save
+ * the ro_image(RX) for calculating the offsets in the
image. The RW
+ * image will be later copied to the RX image from where
the program
+ * will run. The bpf_jit_binary_pack_finalize() will do
this copy in the

arch/riscv/net/bpf_jit_core.c +40 -10
+             * final step.
+             */
+            ctx->ro_insns = (u16 *)jit_data->ro_image;
                ctx->insns = (u16 *)jit_data->image;
                /*
                 * Now, when the image is allocated, the image can
@@ -138,14 +151,12 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
        if (i == NR_JIT_ITERATIONS) {
            pr_err("bpf-jit: image did not converge in <%d passes!\n", i);
-        if (jit_data->header)
-            bpf_jit_binary_free(jit_data->header);
            prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
        }
        if (extable_size)
-        prog->aux->extable = (void *)ctx->insns + prog_size;
+        prog->aux->extable = (void *)ctx->ro_insns + prog_size;
    skip_init_ctx:
        pass++;
@@ -154,23 +165,35 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
        bpf_jit_build_prologue(ctx);
        if (build_body(ctx, extra_pass, NULL)) {
-        bpf_jit_binary_free(jit_data->header);
            prog = orig_prog;
-        goto out_offset;
+        goto out_free_hdr;
        }
        bpf_jit_build_epilogue(ctx);
        if (bpf_jit_enable > 1)
            bpf_jit_dump(prog->len, prog_size, pass, ctx->insns);
-    prog->bpf_func = (void *)ctx->insns;
+    prog->bpf_func = (void *)ctx->ro_insns;
        prog->jited = 1;
        prog->jited_len = prog_size;
-    bpf_flush_icache(jit_data->header, ctx->insns + ctx->ninsns);
-
        if (!prog->is_func || extra_pass) {
-        bpf_jit_binary_lock_ro(jit_data->header);
+        if (WARN_ON(bpf_jit_binary_pack_finalize(prog,
+                             jit_data->ro_header,
+                             jit_data->header))) {
+            /* ro_header has been freed */
+            jit_data->ro_header = NULL;
+            prog = orig_prog;
+            goto out_offset;
+        }
+        /*
+         * The instructions have now been copied to the ROX region from
+         * where they will execute.
+         * Write any modified data cache blocks out to memory and
+         * invalidate the corresponding blocks in the instruction cache.
+         */
+        bpf_flush_icache(jit_data->ro_header,
+                 ctx->ro_insns + ctx->ninsns);
            for (i = 0; i < prog->len; i++)
                ctx->offset[i] = ninsns_rvoff(ctx->offset[i]);
            bpf_prog_fill_jited_linfo(prog, ctx->offset);
@@ -185,6 +208,15 @@ struct bpf_prog *bpf_int_jit_compile(struct
 bpf_prog *prog)
            bpf_jit_prog_release_other(prog, prog == orig_prog ?
                           tmp : orig_prog);
        return prog;
+
+out_free_hdr:
+    if (jit_data->header) {
+        bpf_arch_text_copy(&jit_data->ro_header->size,
+                   &jit_data->header->size,
+                   sizeof(jit_data->header->size));
+        bpf_jit_binary_pack_free(jit_data->ro_header, jit_data->header);
+    }
+    goto out_offset;
    }
    u64 bpf_jit_alloc_exec_limit(void)
@@ -204,3 +236,52 @@ void bpf_jit_free_exec(void *addr)
    {
        return vfree(addr);
    }
+
+void *bpf_arch_text_copy(void *dst, void *src, size_t len)
+{
+    int ret;
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_nosync(dst, src, len);
+    mutex_unlock(&text_mutex);

arch/riscv/net/bpf_jit_core.c void bpf_jit_free_exec(void *addr) · line 256

Quoted 10 lines
arch/riscv/net/bpf_jit_core.c +10 -0
+
+    if (ret)
+        return ERR_PTR(-EINVAL);
+
+    return dst;
+}
+
+int bpf_arch_text_invalidate(void *dst, size_t len)
+{
+    int ret = 0;

no need to initialize it

Quoted 31 lines
arch/riscv/net/bpf_jit_core.c +31 -0
+
+    mutex_lock(&text_mutex);
+    ret = patch_text_set_nosync(dst, 0, len);
+    mutex_unlock(&text_mutex);
+
+    return ret;
+}
+
+void bpf_jit_free(struct bpf_prog *prog)
+{
+    if (prog->jited) {
+        struct rv_jit_data *jit_data = prog->aux->jit_data;
+        struct bpf_binary_header *hdr;
+
+        /*
+         * If we fail the final pass of JIT (from jit_subprogs),
+         * the program may not be finalized yet. Call finalize here
+         * before freeing it.
+         */
+        if (jit_data) {
+            bpf_jit_binary_pack_finalize(prog, jit_data->ro_header,
+                             jit_data->header);
+            kfree(jit_data);
+        }
+        hdr = bpf_jit_binary_pack_hdr(prog);
+        bpf_jit_binary_pack_free(hdr, NULL);
+        WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
+    }
+
+    bpf_prog_unlock_free(prog);
+}
Quoted 2 lines
Quoted 2 lines

Thanks,
Puranjay

Thanks,
Puranjay

Signature · 5 lines
Thanks and Regards

Yours Truly,

Puranjay Mohan
maintainer
#18 lore raw

Re: [PATCH bpf-next v2 3/3] bpf, riscv: use prog pack allocator in the BPF JIT

In reply to Puranjay Mohan
Quoted 8 lines · Puranjay Mohan

Use bpf_jit_binary_pack_alloc() for memory management of JIT binaries in RISCV BPF JIT. The bpf_jit_binary_pack_alloc creates a pair of RW and RX buffers. The JIT writes the program into the RW buffer. When the JIT is done, the program is copied to the final RX buffer with bpf_jit_binary_pack_finalize.

Implement bpf_arch_text_copy() and bpf_arch_text_invalidate() for RISCV JIT as these functions are required by bpf_jit_binary_pack allocator.

General style comment; Please try to use the full 100 characters width for the patches. You're having a lot of linebreaks, which IMO makes the patch harder to read.

Björn

maintainer
#19 lore raw
In reply to Puranjay Mohan
Quoted 28 lines · Puranjay Mohan

Changes in v1 -> v2:
1. Implement a new function patch_text_set_nosync() to be used in bpf_arch_text_invalidate(). The implementation in v1 called patch_text_nosync() in a loop and it was bad as it would call flush_icache_range() for every word making it really slow. This was found by running the test_tag selftest which would take forever to complete.

Here is some data to prove the V2 fixes the problem:

Without this series:

root@rv-selftester:~/src/kselftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m47.562s
user    0m24.145s
sys     6m37.064s

With this series applied:

root@rv-selftester:~/src/selftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m29.472s
user    0m25.865s
sys     6m18.401s

BPF programs currently consume a page each on RISCV. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.

Quoted 8 lines

Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the x86_64 BPF JIT.

I enabled this allocator on the ARM64 BPF JIT[2]. It is being reviewed now.

This patch series enables the BPF prog pack allocator for the RISCV BPF JIT. This series needs a patch[3] from the ARM64 series to work.

Is there a new version for arm64 currently? Maybe we could submit this patch first as a separate patch to avoid dependencies.

Quoted 100 lines

======================================================
Performance Analysis of prog pack allocator on RISCV64
======================================================

Test setup:
===========

Host machine: Debian GNU/Linux 11 (bullseye)
Qemu Version: QEMU emulator version 8.0.3 (Debian 1:8.0.3+dfsg-1) u-boot-qemu Version: 2023.07+dfsg-1
opensbi Version: 1.3-1

To test the performance of the BPF prog pack allocator on RV, a stresser tool[4] linked below was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.

The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered. The script is passed a command which would be run in the above environment.

The script was run with following perf command:

./run.sh "perf stat -a \
         -e iTLB-load-misses \
         -e dTLB-load-misses  \
         -e dTLB-store-misses \
         -e instructions \
         --timeout 60000"

The output of the above command is discussed below before and after enabling the BPF prog pack allocator.

The tests were run on qemu-system-riscv64 with 8 cpus, 16G memory. The rootfs was created using Bjorn's riscv-cross-builder[5] docker container linked below.

Results
=======

Before enabling prog pack allocator:
------------------------------------

Performance counter stats for 'system wide':

            4939048      iTLB-load-misses
            5468689      dTLB-load-misses
             465234      dTLB-store-misses
      1441082097998      instructions
       60.045791200 seconds time elapsed

After enabling prog pack allocator:
-----------------------------------

Performance counter stats for 'system wide':

            3430035      iTLB-load-misses
            5008745      dTLB-load-misses
             409944      dTLB-store-misses
      1441535637988      instructions
       60.046296600 seconds time elapsed

Improvements in metrics
=======================

It was expected that the iTLB-load-misses would decrease as now a single huge page is used to keep all the BPF programs compared to a single page for each program earlier.

--------------------------------------------
The improvement in iTLB-load-misses: -30.5 %
--------------------------------------------

I repeated this expriment more than 100 times in different setups and the improvement was always greater than 30%.

This patch series is boot tested on the Starfive VisionFive 2 board[6]. The performance analysis was not done on the board because it doesn't expose iTLB-load-misses, etc. The stresser program was run on the board to test the loading and unloading of BPF programs

[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/
[4] https://github.com/puranjaymohan/BPF-Allocator-Bench
[5] https://github.com/bjoto/riscv-cross-builder
[6] https://www.starfivetech.com/en/site/boards

Puranjay Mohan (3):
riscv: extend patch_text_nosync() for multiple pages
riscv: implement a memset like function for text bpf, riscv: use prog pack allocator in the BPF JIT

  arch/riscv/include/asm/patch.h  |   1 +
  arch/riscv/kernel/patch.c       | 113 ++++++++++++++++++++++++++++++--
  arch/riscv/net/bpf_jit.h        |   3 +
  arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
  arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
  5 files changed, 255 insertions(+), 31 deletions(-)
#20 lore raw
Quoted 36 lines · Pu Lehui
Quoted 32 lines · Puranjay Mohan

Changes in v1 -> v2:
1. Implement a new function patch_text_set_nosync() to be used in bpf_arch_text_invalidate(). The implementation in v1 called patch_text_nosync() in a loop and it was bad as it would call flush_icache_range() for every word making it really slow. This was found by running the test_tag selftest which would take forever to complete.

Here is some data to prove the V2 fixes the problem:

Without this series:

root@rv-selftester:~/src/kselftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m47.562s
user    0m24.145s
sys     6m37.064s

With this series applied:

root@rv-selftester:~/src/selftest/bpf# time ./test_tag
test_tag: OK (40945 tests)
real    7m29.472s
user    0m25.865s
sys     6m18.401s

BPF programs currently consume a page each on RISCV. For systems with many BPF programs, this adds significant pressure to instruction TLB. High iTLB pressure usually causes slow down for the whole system.

Song Liu introduced the BPF prog pack allocator[1] to mitigate the above issue. It packs multiple BPF programs into a single huge page. It is currently only enabled for the x86_64 BPF JIT.

Quoted 7 lines
Quoted 4 lines

I enabled this allocator on the ARM64 BPF JIT[2]. It is being reviewed now.

This patch series enables the BPF prog pack allocator for the RISCV BPF JIT. This series needs a patch[3] from the ARM64 series to work.

Is there a new version for arm64 currently? Maybe we could submit this patch first as a separate patch to avoid dependencies.

Okay, I will send that patch as a separate patch because it is needed for all architectures.

Quoted 101 lines
Quoted 101 lines

======================================================
Performance Analysis of prog pack allocator on RISCV64
======================================================

Test setup:
===========

Host machine: Debian GNU/Linux 11 (bullseye)
Qemu Version: QEMU emulator version 8.0.3 (Debian 1:8.0.3+dfsg-1) u-boot-qemu Version: 2023.07+dfsg-1
opensbi Version: 1.3-1

To test the performance of the BPF prog pack allocator on RV, a stresser tool[4] linked below was built. This tool loads 8 BPF programs on the system and triggers 5 of them in an infinite loop by doing system calls.

The runner script starts 20 instances of the above which loads 8*20=160 BPF programs on the system, 5*20=100 of which are being constantly triggered. The script is passed a command which would be run in the above environment.

The script was run with following perf command:

./run.sh "perf stat -a \
         -e iTLB-load-misses \
         -e dTLB-load-misses  \
         -e dTLB-store-misses \
         -e instructions \
         --timeout 60000"

The output of the above command is discussed below before and after enabling the BPF prog pack allocator.

The tests were run on qemu-system-riscv64 with 8 cpus, 16G memory. The rootfs was created using Bjorn's riscv-cross-builder[5] docker container linked below.

Results
=======

Before enabling prog pack allocator:
------------------------------------

Performance counter stats for 'system wide':

            4939048      iTLB-load-misses
            5468689      dTLB-load-misses
             465234      dTLB-store-misses
      1441082097998      instructions
       60.045791200 seconds time elapsed

After enabling prog pack allocator:
-----------------------------------

Performance counter stats for 'system wide':

            3430035      iTLB-load-misses
            5008745      dTLB-load-misses
             409944      dTLB-store-misses
      1441535637988      instructions
       60.046296600 seconds time elapsed

Improvements in metrics
=======================

It was expected that the iTLB-load-misses would decrease as now a single huge page is used to keep all the BPF programs compared to a single page for each program earlier.

--------------------------------------------
The improvement in iTLB-load-misses: -30.5 %
--------------------------------------------

I repeated this expriment more than 100 times in different setups and the improvement was always greater than 30%.

This patch series is boot tested on the Starfive VisionFive 2 board[6]. The performance analysis was not done on the board because it doesn't expose iTLB-load-misses, etc. The stresser program was run on the board to test the loading and unloading of BPF programs

[1] https://lore.kernel.org/bpf/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/
[4] https://github.com/puranjaymohan/BPF-Allocator-Bench
[5] https://github.com/bjoto/riscv-cross-builder
[6] https://www.starfivetech.com/en/site/boards

Puranjay Mohan (3):
riscv: extend patch_text_nosync() for multiple pages
riscv: implement a memset like function for text bpf, riscv: use prog pack allocator in the BPF JIT

  arch/riscv/include/asm/patch.h  |   1 +
  arch/riscv/kernel/patch.c       | 113 ++++++++++++++++++++++++++++++--
  arch/riscv/net/bpf_jit.h        |   3 +
  arch/riscv/net/bpf_jit_comp64.c |  56 +++++++++++++---
  arch/riscv/net/bpf_jit_core.c   | 113 +++++++++++++++++++++++++++-----
  5 files changed, 255 insertions(+), 31 deletions(-)
Signature · 5 lines
Thanks and Regards

Yours Truly,

Puranjay Mohan