bpf/verifier: rework value tracking
authorEdward Cree <ecree@solarflare.com>
Mon, 7 Aug 2017 14:26:19 +0000 (15:26 +0100)
committerDavid S. Miller <davem@davemloft.net>
Wed, 9 Aug 2017 00:51:34 +0000 (17:51 -0700)
Unifies adjusted and unadjusted register value types (e.g. FRAME_POINTER is
 now just a PTR_TO_STACK with zero offset).
Tracks value alignment by means of tracking known & unknown bits.  This
 also replaces the 'reg->imm' (leading zero bits) calculations for (what
 were) UNKNOWN_VALUEs.
If pointer leaks are allowed, and adjust_ptr_min_max_vals returns -EACCES,
 treat the pointer as an unknown scalar and try again, because we might be
 able to conclude something about the result (e.g. pointer & 0x40 is either
 0 or 0x40).
Verifier hooks in the netronome/nfp driver were changed to match the new
 data structures.

Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/netronome/nfp/bpf/verifier.c
include/linux/bpf.h
include/linux/bpf_verifier.h
include/linux/tnum.h [new file with mode: 0644]
kernel/bpf/Makefile
kernel/bpf/tnum.c [new file with mode: 0644]
kernel/bpf/verifier.c

index d696ba46f70a6511bea71764cf1dd9578c287929..5b783a91b115e332dbae8447f32e84c1957b8f7d 100644 (file)
@@ -79,28 +79,32 @@ nfp_bpf_check_exit(struct nfp_prog *nfp_prog,
                   const struct bpf_verifier_env *env)
 {
        const struct bpf_reg_state *reg0 = &env->cur_state.regs[0];
+       u64 imm;
 
        if (nfp_prog->act == NN_ACT_XDP)
                return 0;
 
-       if (reg0->type != CONST_IMM) {
-               pr_info("unsupported exit state: %d, imm: %llx\n",
-                       reg0->type, reg0->imm);
+       if (!(reg0->type == SCALAR_VALUE && tnum_is_const(reg0->var_off))) {
+               char tn_buf[48];
+
+               tnum_strn(tn_buf, sizeof(tn_buf), reg0->var_off);
+               pr_info("unsupported exit state: %d, var_off: %s\n",
+                       reg0->type, tn_buf);
                return -EINVAL;
        }
 
-       if (nfp_prog->act != NN_ACT_DIRECT &&
-           reg0->imm != 0 && (reg0->imm & ~0U) != ~0U) {
+       imm = reg0->var_off.value;
+       if (nfp_prog->act != NN_ACT_DIRECT && imm != 0 && (imm & ~0U) != ~0U) {
                pr_info("unsupported exit state: %d, imm: %llx\n",
-                       reg0->type, reg0->imm);
+                       reg0->type, imm);
                return -EINVAL;
        }
 
-       if (nfp_prog->act == NN_ACT_DIRECT && reg0->imm <= TC_ACT_REDIRECT &&
-           reg0->imm != TC_ACT_SHOT && reg0->imm != TC_ACT_STOLEN &&
-           reg0->imm != TC_ACT_QUEUED) {
+       if (nfp_prog->act == NN_ACT_DIRECT && imm <= TC_ACT_REDIRECT &&
+           imm != TC_ACT_SHOT && imm != TC_ACT_STOLEN &&
+           imm != TC_ACT_QUEUED) {
                pr_info("unsupported exit state: %d, imm: %llx\n",
-                       reg0->type, reg0->imm);
+                       reg0->type, imm);
                return -EINVAL;
        }
 
index 6353c7474dba6ba590b4f9856aea5a395d0741a4..39229c455cba945512a46f902cda0d319d56d8c3 100644 (file)
@@ -117,35 +117,25 @@ enum bpf_access_type {
 };
 
 /* types of values stored in eBPF registers */
+/* Pointer types represent:
+ * pointer
+ * pointer + imm
+ * pointer + (u16) var
+ * pointer + (u16) var + imm
+ * if (range > 0) then [ptr, ptr + range - off) is safe to access
+ * if (id > 0) means that some 'var' was added
+ * if (off > 0) means that 'imm' was added
+ */
 enum bpf_reg_type {
        NOT_INIT = 0,            /* nothing was written into register */
-       UNKNOWN_VALUE,           /* reg doesn't contain a valid pointer */
+       SCALAR_VALUE,            /* reg doesn't contain a valid pointer */
        PTR_TO_CTX,              /* reg points to bpf_context */
        CONST_PTR_TO_MAP,        /* reg points to struct bpf_map */
        PTR_TO_MAP_VALUE,        /* reg points to map element value */
        PTR_TO_MAP_VALUE_OR_NULL,/* points to map elem value or NULL */
-       FRAME_PTR,               /* reg == frame_pointer */
-       PTR_TO_STACK,            /* reg == frame_pointer + imm */
-       CONST_IMM,               /* constant integer value */
-
-       /* PTR_TO_PACKET represents:
-        * skb->data
-        * skb->data + imm
-        * skb->data + (u16) var
-        * skb->data + (u16) var + imm
-        * if (range > 0) then [ptr, ptr + range - off) is safe to access
-        * if (id > 0) means that some 'var' was added
-        * if (off > 0) menas that 'imm' was added
-        */
-       PTR_TO_PACKET,
+       PTR_TO_STACK,            /* reg == frame_pointer + offset */
+       PTR_TO_PACKET,           /* reg points to skb->data */
        PTR_TO_PACKET_END,       /* skb->data + headlen */
-
-       /* PTR_TO_MAP_VALUE_ADJ is used for doing pointer math inside of a map
-        * elem value.  We only allow this if we can statically verify that
-        * access from this register are going to fall within the size of the
-        * map element.
-        */
-       PTR_TO_MAP_VALUE_ADJ,
 };
 
 struct bpf_prog;
index 8e5d31f6faefd361cee145be654a968123ae5e6a..85936fa92d122f8dae6e36ac837d5333b78483af 100644 (file)
@@ -9,6 +9,7 @@
 
 #include <linux/bpf.h> /* for enum bpf_reg_type */
 #include <linux/filter.h> /* for MAX_BPF_STACK */
+#include <linux/tnum.h>
 
  /* Just some arbitrary values so we can safely do math without overflowing and
   * are obviously wrong for any sort of memory access.
 struct bpf_reg_state {
        enum bpf_reg_type type;
        union {
-               /* valid when type == CONST_IMM | PTR_TO_STACK | UNKNOWN_VALUE */
-               s64 imm;
-
-               /* valid when type == PTR_TO_PACKET* */
-               struct {
-                       u16 off;
-                       u16 range;
-               };
+               /* valid when type == PTR_TO_PACKET */
+               u16 range;
 
                /* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE |
                 *   PTR_TO_MAP_VALUE_OR_NULL
                 */
                struct bpf_map *map_ptr;
        };
+       /* Fixed part of pointer offset, pointer types only */
+       s32 off;
+       /* For PTR_TO_PACKET, used to find other pointers with the same variable
+        * offset, so they can share range knowledge.
+        * For PTR_TO_MAP_VALUE_OR_NULL this is used to share which map value we
+        * came from, when one is tested for != NULL.
+        */
        u32 id;
+       /* These three fields must be last.  See states_equal() */
+       /* For scalar types (SCALAR_VALUE), this represents our knowledge of
+        * the actual value.
+        * For pointer types, this represents the variable part of the offset
+        * from the pointed-to object, and is shared with all bpf_reg_states
+        * with the same id as us.
+        */
+       struct tnum var_off;
        /* Used to determine if any memory access using this register will
-        * result in a bad access. These two fields must be last.
-        * See states_equal()
+        * result in a bad access.
+        * These refer to the same value as var_off, not necessarily the actual
+        * contents of the register.
         */
        s64 min_value;
        u64 max_value;
-       u32 min_align;
-       u32 aux_off;
-       u32 aux_off_align;
        bool value_from_signed;
 };
 
diff --git a/include/linux/tnum.h b/include/linux/tnum.h
new file mode 100644 (file)
index 0000000..a0b07bf
--- /dev/null
@@ -0,0 +1,79 @@
+/* tnum: tracked (or tristate) numbers
+ *
+ * A tnum tracks knowledge about the bits of a value.  Each bit can be either
+ * known (0 or 1), or unknown (x).  Arithmetic operations on tnums will
+ * propagate the unknown bits such that the tnum result represents all the
+ * possible results for possible values of the operands.
+ */
+#include <linux/types.h>
+
+struct tnum {
+       u64 value;
+       u64 mask;
+};
+
+/* Constructors */
+/* Represent a known constant as a tnum. */
+struct tnum tnum_const(u64 value);
+/* A completely unknown value */
+extern const struct tnum tnum_unknown;
+
+/* Arithmetic and logical ops */
+/* Shift a tnum left (by a fixed shift) */
+struct tnum tnum_lshift(struct tnum a, u8 shift);
+/* Shift a tnum right (by a fixed shift) */
+struct tnum tnum_rshift(struct tnum a, u8 shift);
+/* Add two tnums, return @a + @b */
+struct tnum tnum_add(struct tnum a, struct tnum b);
+/* Subtract two tnums, return @a - @b */
+struct tnum tnum_sub(struct tnum a, struct tnum b);
+/* Bitwise-AND, return @a & @b */
+struct tnum tnum_and(struct tnum a, struct tnum b);
+/* Bitwise-OR, return @a | @b */
+struct tnum tnum_or(struct tnum a, struct tnum b);
+/* Bitwise-XOR, return @a ^ @b */
+struct tnum tnum_xor(struct tnum a, struct tnum b);
+/* Multiply two tnums, return @a * @b */
+struct tnum tnum_mul(struct tnum a, struct tnum b);
+
+/* Return a tnum representing numbers satisfying both @a and @b */
+struct tnum tnum_intersect(struct tnum a, struct tnum b);
+
+/* Return @a with all but the lowest @size bytes cleared */
+struct tnum tnum_cast(struct tnum a, u8 size);
+
+/* Returns true if @a is a known constant */
+static inline bool tnum_is_const(struct tnum a)
+{
+       return !a.mask;
+}
+
+/* Returns true if @a == tnum_const(@b) */
+static inline bool tnum_equals_const(struct tnum a, u64 b)
+{
+       return tnum_is_const(a) && a.value == b;
+}
+
+/* Returns true if @a is completely unknown */
+static inline bool tnum_is_unknown(struct tnum a)
+{
+       return !~a.mask;
+}
+
+/* Returns true if @a is known to be a multiple of @size.
+ * @size must be a power of two.
+ */
+bool tnum_is_aligned(struct tnum a, u64 size);
+
+/* Returns true if @b represents a subset of @a. */
+bool tnum_in(struct tnum a, struct tnum b);
+
+/* Formatting functions.  These have snprintf-like semantics: they will write
+ * up to @size bytes (including the terminating NUL byte), and return the number
+ * of bytes (excluding the terminating NUL) which would have been written had
+ * sufficient space been available.  (Thus tnum_sbin always returns 64.)
+ */
+/* Format a tnum as a pair of hex numbers (value; mask) */
+int tnum_strn(char *str, size_t size, struct tnum a);
+/* Format a tnum as tristate binary expansion */
+int tnum_sbin(char *str, size_t size, struct tnum a);
index 48e92705be5976df04c24de687288db537a135b9..2f0bcda40e902d5befec49944eae84932b81f663 100644 (file)
@@ -1,6 +1,6 @@
 obj-y := core.o
 
-obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o
+obj-$(CONFIG_BPF_SYSCALL) += syscall.o verifier.o inode.o helpers.o tnum.o
 obj-$(CONFIG_BPF_SYSCALL) += hashtab.o arraymap.o percpu_freelist.o bpf_lru_list.o lpm_trie.o map_in_map.o
 ifeq ($(CONFIG_NET),y)
 obj-$(CONFIG_BPF_SYSCALL) += devmap.o
diff --git a/kernel/bpf/tnum.c b/kernel/bpf/tnum.c
new file mode 100644 (file)
index 0000000..92eeeb1
--- /dev/null
@@ -0,0 +1,164 @@
+/* tnum: tracked (or tristate) numbers
+ *
+ * A tnum tracks knowledge about the bits of a value.  Each bit can be either
+ * known (0 or 1), or unknown (x).  Arithmetic operations on tnums will
+ * propagate the unknown bits such that the tnum result represents all the
+ * possible results for possible values of the operands.
+ */
+#include <linux/kernel.h>
+#include <linux/tnum.h>
+
+#define TNUM(_v, _m)   (struct tnum){.value = _v, .mask = _m}
+/* A completely unknown value */
+const struct tnum tnum_unknown = { .value = 0, .mask = -1 };
+
+struct tnum tnum_const(u64 value)
+{
+       return TNUM(value, 0);
+}
+
+struct tnum tnum_lshift(struct tnum a, u8 shift)
+{
+       return TNUM(a.value << shift, a.mask << shift);
+}
+
+struct tnum tnum_rshift(struct tnum a, u8 shift)
+{
+       return TNUM(a.value >> shift, a.mask >> shift);
+}
+
+struct tnum tnum_add(struct tnum a, struct tnum b)
+{
+       u64 sm, sv, sigma, chi, mu;
+
+       sm = a.mask + b.mask;
+       sv = a.value + b.value;
+       sigma = sm + sv;
+       chi = sigma ^ sv;
+       mu = chi | a.mask | b.mask;
+       return TNUM(sv & ~mu, mu);
+}
+
+struct tnum tnum_sub(struct tnum a, struct tnum b)
+{
+       u64 dv, alpha, beta, chi, mu;
+
+       dv = a.value - b.value;
+       alpha = dv + a.mask;
+       beta = dv - b.mask;
+       chi = alpha ^ beta;
+       mu = chi | a.mask | b.mask;
+       return TNUM(dv & ~mu, mu);
+}
+
+struct tnum tnum_and(struct tnum a, struct tnum b)
+{
+       u64 alpha, beta, v;
+
+       alpha = a.value | a.mask;
+       beta = b.value | b.mask;
+       v = a.value & b.value;
+       return TNUM(v, alpha & beta & ~v);
+}
+
+struct tnum tnum_or(struct tnum a, struct tnum b)
+{
+       u64 v, mu;
+
+       v = a.value | b.value;
+       mu = a.mask | b.mask;
+       return TNUM(v, mu & ~v);
+}
+
+struct tnum tnum_xor(struct tnum a, struct tnum b)
+{
+       u64 v, mu;
+
+       v = a.value ^ b.value;
+       mu = a.mask | b.mask;
+       return TNUM(v & ~mu, mu);
+}
+
+/* half-multiply add: acc += (unknown * mask * value).
+ * An intermediate step in the multiply algorithm.
+ */
+static struct tnum hma(struct tnum acc, u64 value, u64 mask)
+{
+       while (mask) {
+               if (mask & 1)
+                       acc = tnum_add(acc, TNUM(0, value));
+               mask >>= 1;
+               value <<= 1;
+       }
+       return acc;
+}
+
+struct tnum tnum_mul(struct tnum a, struct tnum b)
+{
+       struct tnum acc;
+       u64 pi;
+
+       pi = a.value * b.value;
+       acc = hma(TNUM(pi, 0), a.mask, b.mask | b.value);
+       return hma(acc, b.mask, a.value);
+}
+
+/* Note that if a and b disagree - i.e. one has a 'known 1' where the other has
+ * a 'known 0' - this will return a 'known 1' for that bit.
+ */
+struct tnum tnum_intersect(struct tnum a, struct tnum b)
+{
+       u64 v, mu;
+
+       v = a.value | b.value;
+       mu = a.mask & b.mask;
+       return TNUM(v & ~mu, mu);
+}
+
+struct tnum tnum_cast(struct tnum a, u8 size)
+{
+       a.value &= (1ULL << (size * 8)) - 1;
+       a.mask &= (1ULL << (size * 8)) - 1;
+       return a;
+}
+
+bool tnum_is_aligned(struct tnum a, u64 size)
+{
+       if (!size)
+               return true;
+       return !((a.value | a.mask) & (size - 1));
+}
+
+bool tnum_in(struct tnum a, struct tnum b)
+{
+       if (b.mask & ~a.mask)
+               return false;
+       b.value &= ~a.mask;
+       return a.value == b.value;
+}
+
+int tnum_strn(char *str, size_t size, struct tnum a)
+{
+       return snprintf(str, size, "(%#llx; %#llx)", a.value, a.mask);
+}
+EXPORT_SYMBOL_GPL(tnum_strn);
+
+int tnum_sbin(char *str, size_t size, struct tnum a)
+{
+       size_t n;
+
+       for (n = 64; n; n--) {
+               if (n < size) {
+                       if (a.mask & 1)
+                               str[n - 1] = 'x';
+                       else if (a.value & 1)
+                               str[n - 1] = '1';
+                       else
+                               str[n - 1] = '0';
+               }
+               a.mask >>= 1;
+               a.value >>= 1;
+       }
+       str[min(size - 1, (size_t)64)] = 0;
+       return 64;
+}
index f6e8b3887eab4862b9db6f366a652bdb719417de..c3f88b466c3049c67991922904fe74783ff93f18 100644 (file)
  * (and -20 constant is saved for further stack bounds checking).
  * Meaning that this reg is a pointer to stack plus known immediate constant.
  *
- * Most of the time the registers have UNKNOWN_VALUE type, which
+ * Most of the time the registers have SCALAR_VALUE type, which
  * means the register has some value, but it's not a valid pointer.
- * (like pointer plus pointer becomes UNKNOWN_VALUE type)
+ * (like pointer plus pointer becomes SCALAR_VALUE type)
  *
  * When verifier sees load or store instructions the type of base register
- * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, FRAME_PTR. These are three pointer
+ * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK. These are three pointer
  * types recognized by check_mem_access() function.
  *
  * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
@@ -180,15 +180,12 @@ static __printf(1, 2) void verbose(const char *fmt, ...)
 /* string representation of 'enum bpf_reg_type' */
 static const char * const reg_type_str[] = {
        [NOT_INIT]              = "?",
-       [UNKNOWN_VALUE]         = "inv",
+       [SCALAR_VALUE]          = "inv",
        [PTR_TO_CTX]            = "ctx",
        [CONST_PTR_TO_MAP]      = "map_ptr",
        [PTR_TO_MAP_VALUE]      = "map_value",
        [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
-       [PTR_TO_MAP_VALUE_ADJ]  = "map_value_adj",
-       [FRAME_PTR]             = "fp",
        [PTR_TO_STACK]          = "fp",
-       [CONST_IMM]             = "imm",
        [PTR_TO_PACKET]         = "pkt",
        [PTR_TO_PACKET_END]     = "pkt_end",
 };
@@ -221,32 +218,36 @@ static void print_verifier_state(struct bpf_verifier_state *state)
                if (t == NOT_INIT)
                        continue;
                verbose(" R%d=%s", i, reg_type_str[t]);
-               if (t == CONST_IMM || t == PTR_TO_STACK)
-                       verbose("%lld", reg->imm);
-               else if (t == PTR_TO_PACKET)
-                       verbose("(id=%d,off=%d,r=%d)",
-                               reg->id, reg->off, reg->range);
-               else if (t == UNKNOWN_VALUE && reg->imm)
-                       verbose("%lld", reg->imm);
-               else if (t == CONST_PTR_TO_MAP || t == PTR_TO_MAP_VALUE ||
-                        t == PTR_TO_MAP_VALUE_OR_NULL ||
-                        t == PTR_TO_MAP_VALUE_ADJ)
-                       verbose("(ks=%d,vs=%d,id=%u)",
-                               reg->map_ptr->key_size,
-                               reg->map_ptr->value_size,
-                               reg->id);
-               if (reg->min_value != BPF_REGISTER_MIN_RANGE)
-                       verbose(",min_value=%lld",
-                               (long long)reg->min_value);
-               if (reg->max_value != BPF_REGISTER_MAX_RANGE)
-                       verbose(",max_value=%llu",
-                               (unsigned long long)reg->max_value);
-               if (reg->min_align)
-                       verbose(",min_align=%u", reg->min_align);
-               if (reg->aux_off)
-                       verbose(",aux_off=%u", reg->aux_off);
-               if (reg->aux_off_align)
-                       verbose(",aux_off_align=%u", reg->aux_off_align);
+               if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
+                   tnum_is_const(reg->var_off)) {
+                       /* reg->off should be 0 for SCALAR_VALUE */
+                       verbose("%lld", reg->var_off.value + reg->off);
+               } else {
+                       verbose("(id=%d", reg->id);
+                       if (t != SCALAR_VALUE)
+                               verbose(",off=%d", reg->off);
+                       if (t == PTR_TO_PACKET)
+                               verbose(",r=%d", reg->range);
+                       else if (t == CONST_PTR_TO_MAP ||
+                                t == PTR_TO_MAP_VALUE ||
+                                t == PTR_TO_MAP_VALUE_OR_NULL)
+                               verbose(",ks=%d,vs=%d",
+                                       reg->map_ptr->key_size,
+                                       reg->map_ptr->value_size);
+                       if (reg->min_value != BPF_REGISTER_MIN_RANGE)
+                               verbose(",min_value=%lld",
+                                       (long long)reg->min_value);
+                       if (reg->max_value != BPF_REGISTER_MAX_RANGE)
+                               verbose(",max_value=%llu",
+                                       (unsigned long long)reg->max_value);
+                       if (!tnum_is_unknown(reg->var_off)) {
+                               char tn_buf[48];
+
+                               tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
+                               verbose(",var_off=%s", tn_buf);
+                       }
+                       verbose(")");
+               }
        }
        for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
                if (state->stack_slot_type[i] == STACK_SPILL)
@@ -463,14 +464,69 @@ static const int caller_saved[CALLER_SAVED_REGS] = {
        BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
 };
 
-static void mark_reg_not_init(struct bpf_reg_state *regs, u32 regno)
+static void __mark_reg_not_init(struct bpf_reg_state *reg);
+
+/* Mark the 'variable offset' part of a register as zero.  This should be
+ * used only on registers holding a pointer type.
+ */
+static void __mark_reg_known_zero(struct bpf_reg_state *reg)
 {
-       BUG_ON(regno >= MAX_BPF_REG);
+       reg->var_off = tnum_const(0);
+       reg->min_value = 0;
+       reg->max_value = 0;
+}
 
-       memset(&regs[regno], 0, sizeof(regs[regno]));
-       regs[regno].type = NOT_INIT;
-       regs[regno].min_value = BPF_REGISTER_MIN_RANGE;
-       regs[regno].max_value = BPF_REGISTER_MAX_RANGE;
+static void mark_reg_known_zero(struct bpf_reg_state *regs, u32 regno)
+{
+       if (WARN_ON(regno >= MAX_BPF_REG)) {
+               verbose("mark_reg_known_zero(regs, %u)\n", regno);
+               /* Something bad happened, let's kill all regs */
+               for (regno = 0; regno < MAX_BPF_REG; regno++)
+                       __mark_reg_not_init(regs + regno);
+               return;
+       }
+       __mark_reg_known_zero(regs + regno);
+}
+
+/* Mark a register as having a completely unknown (scalar) value. */
+static void __mark_reg_unknown(struct bpf_reg_state *reg)
+{
+       reg->type = SCALAR_VALUE;
+       reg->id = 0;
+       reg->off = 0;
+       reg->var_off = tnum_unknown;
+       reg->min_value = BPF_REGISTER_MIN_RANGE;
+       reg->max_value = BPF_REGISTER_MAX_RANGE;
+}
+
+static void mark_reg_unknown(struct bpf_reg_state *regs, u32 regno)
+{
+       if (WARN_ON(regno >= MAX_BPF_REG)) {
+               verbose("mark_reg_unknown(regs, %u)\n", regno);
+               /* Something bad happened, let's kill all regs */
+               for (regno = 0; regno < MAX_BPF_REG; regno++)
+                       __mark_reg_not_init(regs + regno);
+               return;
+       }
+       __mark_reg_unknown(regs + regno);
+}
+
+static void __mark_reg_not_init(struct bpf_reg_state *reg)
+{
+       __mark_reg_unknown(reg);
+       reg->type = NOT_INIT;
+}
+
+static void mark_reg_not_init(struct bpf_reg_state *regs, u32 regno)
+{
+       if (WARN_ON(regno >= MAX_BPF_REG)) {
+               verbose("mark_reg_not_init(regs, %u)\n", regno);
+               /* Something bad happened, let's kill all regs */
+               for (regno = 0; regno < MAX_BPF_REG; regno++)
+                       __mark_reg_not_init(regs + regno);
+               return;
+       }
+       __mark_reg_not_init(regs + regno);
 }
 
 static void init_reg_state(struct bpf_reg_state *regs)
@@ -481,23 +537,12 @@ static void init_reg_state(struct bpf_reg_state *regs)
                mark_reg_not_init(regs, i);
 
        /* frame pointer */
-       regs[BPF_REG_FP].type = FRAME_PTR;
+       regs[BPF_REG_FP].type = PTR_TO_STACK;
+       mark_reg_known_zero(regs, BPF_REG_FP);
 
        /* 1st arg to a function */
        regs[BPF_REG_1].type = PTR_TO_CTX;
-}
-
-static void __mark_reg_unknown_value(struct bpf_reg_state *regs, u32 regno)
-{
-       regs[regno].type = UNKNOWN_VALUE;
-       regs[regno].id = 0;
-       regs[regno].imm = 0;
-}
-
-static void mark_reg_unknown_value(struct bpf_reg_state *regs, u32 regno)
-{
-       BUG_ON(regno >= MAX_BPF_REG);
-       __mark_reg_unknown_value(regs, regno);
+       mark_reg_known_zero(regs, BPF_REG_1);
 }
 
 static void reset_reg_range_values(struct bpf_reg_state *regs, u32 regno)
@@ -505,14 +550,6 @@ static void reset_reg_range_values(struct bpf_reg_state *regs, u32 regno)
        regs[regno].min_value = BPF_REGISTER_MIN_RANGE;
        regs[regno].max_value = BPF_REGISTER_MAX_RANGE;
        regs[regno].value_from_signed = false;
-       regs[regno].min_align = 0;
-}
-
-static void mark_reg_unknown_value_and_range(struct bpf_reg_state *regs,
-                                            u32 regno)
-{
-       mark_reg_unknown_value(regs, regno);
-       reset_reg_range_values(regs, regno);
 }
 
 enum reg_arg_type {
@@ -542,7 +579,7 @@ static int check_reg_arg(struct bpf_reg_state *regs, u32 regno,
                        return -EACCES;
                }
                if (t == DST_OP)
-                       mark_reg_unknown_value(regs, regno);
+                       mark_reg_unknown(regs, regno);
        }
        return 0;
 }
@@ -552,12 +589,10 @@ static bool is_spillable_regtype(enum bpf_reg_type type)
        switch (type) {
        case PTR_TO_MAP_VALUE:
        case PTR_TO_MAP_VALUE_OR_NULL:
-       case PTR_TO_MAP_VALUE_ADJ:
        case PTR_TO_STACK:
        case PTR_TO_CTX:
        case PTR_TO_PACKET:
        case PTR_TO_PACKET_END:
-       case FRAME_PTR:
        case CONST_PTR_TO_MAP:
                return true;
        default:
@@ -637,14 +672,13 @@ static int check_stack_read(struct bpf_verifier_state *state, int off, int size,
                }
                if (value_regno >= 0)
                        /* have read misc data from the stack */
-                       mark_reg_unknown_value_and_range(state->regs,
-                                                        value_regno);
+                       mark_reg_unknown(state->regs, value_regno);
                return 0;
        }
 }
 
 /* check read/write into map element returned by bpf_map_lookup_elem() */
-static int check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
+static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
                            int size)
 {
        struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
@@ -657,22 +691,25 @@ static int check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
        return 0;
 }
 
-/* check read/write into an adjusted map element */
-static int check_map_access_adj(struct bpf_verifier_env *env, u32 regno,
+/* check read/write into a map element with possible variable offset */
+static int check_map_access(struct bpf_verifier_env *env, u32 regno,
                                int off, int size)
 {
        struct bpf_verifier_state *state = &env->cur_state;
        struct bpf_reg_state *reg = &state->regs[regno];
        int err;
 
-       /* We adjusted the register to this map value, so we
-        * need to change off and size to min_value and max_value
-        * respectively to make sure our theoretical access will be
-        * safe.
+       /* We may have adjusted the register to this map value, so we
+        * need to try adding each of min_value and max_value to off
+        * to make sure our theoretical access will be safe.
         */
        if (log_level)
                print_verifier_state(state);
-       env->varlen_map_value_access = true;
+       /* If the offset is variable, we will need to be stricter in state
+        * pruning from now on.
+        */
+       if (!tnum_is_const(reg->var_off))
+               env->varlen_map_value_access = true;
        /* The minimum value is only important with signed
         * comparisons where we can't assume the floor of a
         * value is 0.  If we are using signed variables for our
@@ -684,10 +721,9 @@ static int check_map_access_adj(struct bpf_verifier_env *env, u32 regno,
                        regno);
                return -EACCES;
        }
-       err = check_map_access(env, regno, reg->min_value + off, size);
+       err = __check_map_access(env, regno, reg->min_value + off, size);
        if (err) {
-               verbose("R%d min value is outside of the array range\n",
-                       regno);
+               verbose("R%d min value is outside of the array range\n", regno);
                return err;
        }
 
@@ -699,7 +735,10 @@ static int check_map_access_adj(struct bpf_verifier_env *env, u32 regno,
                        regno);
                return -EACCES;
        }
-       return check_map_access(env, regno, reg->max_value + off, size);
+       err = __check_map_access(env, regno, reg->max_value + off, size);
+       if (err)
+               verbose("R%d max value is outside of the array range\n", regno);
+       return err;
 }
 
 #define MAX_PACKET_OFF 0xffff
@@ -729,14 +768,13 @@ static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
        }
 }
 
-static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
-                              int size)
+static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
+                                int off, int size)
 {
        struct bpf_reg_state *regs = env->cur_state.regs;
        struct bpf_reg_state *reg = &regs[regno];
 
-       off += reg->off;
-       if (off < 0 || size <= 0 || off + size > reg->range) {
+       if (off < 0 || size <= 0 || (u64)off + size > reg->range) {
                verbose("invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
                        off, size, regno, reg->id, reg->off, reg->range);
                return -EACCES;
@@ -744,7 +782,35 @@ static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
        return 0;
 }
 
-/* check access to 'struct bpf_context' fields */
+static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
+                              int size)
+{
+       struct bpf_reg_state *regs = env->cur_state.regs;
+       struct bpf_reg_state *reg = &regs[regno];
+       int err;
+
+       /* We may have added a variable offset to the packet pointer; but any
+        * reg->range we have comes after that.  We are only checking the fixed
+        * offset.
+        */
+
+       /* We don't allow negative numbers, because we aren't tracking enough
+        * detail to prove they're safe.
+        */
+       if (reg->min_value < 0) {
+               verbose("R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
+                       regno);
+               return -EACCES;
+       }
+       err = __check_packet_access(env, regno, off, size);
+       if (err) {
+               verbose("R%d offset is outside of the packet\n", regno);
+               return err;
+       }
+       return err;
+}
+
+/* check access to 'struct bpf_context' fields.  Supports fixed offsets only */
 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
                            enum bpf_access_type t, enum bpf_reg_type *reg_type)
 {
@@ -784,13 +850,7 @@ static bool __is_pointer_value(bool allow_ptr_leaks,
        if (allow_ptr_leaks)
                return false;
 
-       switch (reg->type) {
-       case UNKNOWN_VALUE:
-       case CONST_IMM:
-               return false;
-       default:
-               return true;
-       }
+       return reg->type != SCALAR_VALUE;
 }
 
 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
@@ -801,23 +861,13 @@ static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
 static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
                                   int off, int size, bool strict)
 {
+       struct tnum reg_off;
        int ip_align;
-       int reg_off;
 
        /* Byte size accesses are always allowed. */
        if (!strict || size == 1)
                return 0;
 
-       reg_off = reg->off;
-       if (reg->id) {
-               if (reg->aux_off_align % size) {
-                       verbose("Packet access is only %u byte aligned, %d byte access not allowed\n",
-                               reg->aux_off_align, size);
-                       return -EACCES;
-               }
-               reg_off += reg->aux_off;
-       }
-
        /* For platforms that do not have a Kconfig enabling
         * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
         * NET_IP_ALIGN is universally set to '2'.  And on platforms
@@ -827,20 +877,37 @@ static int check_pkt_ptr_alignment(const struct bpf_reg_state *reg,
         * unconditional IP align value of '2'.
         */
        ip_align = 2;
-       if ((ip_align + reg_off + off) % size != 0) {
-               verbose("misaligned packet access off %d+%d+%d size %d\n",
-                       ip_align, reg_off, off, size);
+
+       reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
+       if (!tnum_is_aligned(reg_off, size)) {
+               char tn_buf[48];
+
+               tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
+               verbose("misaligned packet access off %d+%s+%d+%d size %d\n",
+                       ip_align, tn_buf, reg->off, off, size);
                return -EACCES;
        }
 
        return 0;
 }
 
-static int check_val_ptr_alignment(const struct bpf_reg_state *reg,
-                                  int size, bool strict)
+static int check_generic_ptr_alignment(const struct bpf_reg_state *reg,
+                                      const char *pointer_desc,
+                                      int off, int size, bool strict)
 {
-       if (strict && size != 1) {
-               verbose("Unknown alignment. Only byte-sized access allowed in value access.\n");
+       struct tnum reg_off;
+
+       /* Byte size accesses are always allowed. */
+       if (!strict || size == 1)
+               return 0;
+
+       reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
+       if (!tnum_is_aligned(reg_off, size)) {
+               char tn_buf[48];
+
+               tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
+               verbose("misaligned %saccess off %s+%d+%d size %d\n",
+                       pointer_desc, tn_buf, reg->off, off, size);
                return -EACCES;
        }
 
@@ -852,21 +919,25 @@ static int check_ptr_alignment(struct bpf_verifier_env *env,
                               int off, int size)
 {
        bool strict = env->strict_alignment;
+       const char *pointer_desc = "";
 
        switch (reg->type) {
        case PTR_TO_PACKET:
+               /* special case, because of NET_IP_ALIGN */
                return check_pkt_ptr_alignment(reg, off, size, strict);
-       case PTR_TO_MAP_VALUE_ADJ:
-               return check_val_ptr_alignment(reg, size, strict);
+       case PTR_TO_MAP_VALUE:
+               pointer_desc = "value ";
+               break;
+       case PTR_TO_CTX:
+               pointer_desc = "context ";
+               break;
+       case PTR_TO_STACK:
+               pointer_desc = "stack ";
+               break;
        default:
-               if (off % size != 0) {
-                       verbose("misaligned access off %d size %d\n",
-                               off, size);
-                       return -EACCES;
-               }
-
-               return 0;
+               break;
        }
+       return check_generic_ptr_alignment(reg, pointer_desc, off, size, strict);
 }
 
 /* check whether memory at (regno + off) is accessible for t = (read | write)
@@ -883,52 +954,79 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
        struct bpf_reg_state *reg = &state->regs[regno];
        int size, err = 0;
 
-       if (reg->type == PTR_TO_STACK)
-               off += reg->imm;
-
        size = bpf_size_to_bytes(bpf_size);
        if (size < 0)
                return size;
 
+       /* alignment checks will add in reg->off themselves */
        err = check_ptr_alignment(env, reg, off, size);
        if (err)
                return err;
 
-       if (reg->type == PTR_TO_MAP_VALUE ||
-           reg->type == PTR_TO_MAP_VALUE_ADJ) {
+       /* for access checks, reg->off is just part of off */
+       off += reg->off;
+
+       if (reg->type == PTR_TO_MAP_VALUE) {
                if (t == BPF_WRITE && value_regno >= 0 &&
                    is_pointer_value(env, value_regno)) {
                        verbose("R%d leaks addr into map\n", value_regno);
                        return -EACCES;
                }
 
-               if (reg->type == PTR_TO_MAP_VALUE_ADJ)
-                       err = check_map_access_adj(env, regno, off, size);
-               else
-                       err = check_map_access(env, regno, off, size);
+               err = check_map_access(env, regno, off, size);
                if (!err && t == BPF_READ && value_regno >= 0)
-                       mark_reg_unknown_value_and_range(state->regs,
-                                                        value_regno);
+                       mark_reg_unknown(state->regs, value_regno);
 
        } else if (reg->type == PTR_TO_CTX) {
-               enum bpf_reg_type reg_type = UNKNOWN_VALUE;
+               enum bpf_reg_type reg_type = SCALAR_VALUE;
 
                if (t == BPF_WRITE && value_regno >= 0 &&
                    is_pointer_value(env, value_regno)) {
                        verbose("R%d leaks addr into ctx\n", value_regno);
                        return -EACCES;
                }
+               /* ctx accesses must be at a fixed offset, so that we can
+                * determine what type of data were returned.
+                */
+               if (!tnum_is_const(reg->var_off)) {
+                       char tn_buf[48];
+
+                       tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
+                       verbose("variable ctx access var_off=%s off=%d size=%d",
+                               tn_buf, off, size);
+                       return -EACCES;
+               }
+               off += reg->var_off.value;
                err = check_ctx_access(env, insn_idx, off, size, t, &reg_type);
                if (!err && t == BPF_READ && value_regno >= 0) {
-                       mark_reg_unknown_value_and_range(state->regs,
-                                                        value_regno);
-                       /* note that reg.[id|off|range] == 0 */
+                       /* ctx access returns either a scalar, or a
+                        * PTR_TO_PACKET[_END].  In the latter case, we know
+                        * the offset is zero.
+                        */
+                       if (reg_type == SCALAR_VALUE)
+                               mark_reg_unknown(state->regs, value_regno);
+                       else
+                               mark_reg_known_zero(state->regs, value_regno);
+                       state->regs[value_regno].id = 0;
+                       state->regs[value_regno].off = 0;
+                       state->regs[value_regno].range = 0;
                        state->regs[value_regno].type = reg_type;
-                       state->regs[value_regno].aux_off = 0;
-                       state->regs[value_regno].aux_off_align = 0;
                }
 
-       } else if (reg->type == FRAME_PTR || reg->type == PTR_TO_STACK) {
+       } else if (reg->type == PTR_TO_STACK) {
+               /* stack accesses must be at a fixed offset, so that we can
+                * determine what type of data were returned.
+                * See check_stack_read().
+                */
+               if (!tnum_is_const(reg->var_off)) {
+                       char tn_buf[48];
+
+                       tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
+                       verbose("variable stack access var_off=%s off=%d size=%d",
+                               tn_buf, off, size);
+                       return -EACCES;
+               }
+               off += reg->var_off.value;
                if (off >= 0 || off < -MAX_BPF_STACK) {
                        verbose("invalid stack off=%d size=%d\n", off, size);
                        return -EACCES;
@@ -948,7 +1046,7 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
                } else {
                        err = check_stack_read(state, off, size, value_regno);
                }
-       } else if (state->regs[regno].type == PTR_TO_PACKET) {
+       } else if (reg->type == PTR_TO_PACKET) {
                if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
                        verbose("cannot write into packet\n");
                        return -EACCES;
@@ -960,21 +1058,24 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
                }
                err = check_packet_access(env, regno, off, size);
                if (!err && t == BPF_READ && value_regno >= 0)
-                       mark_reg_unknown_value_and_range(state->regs,
-                                                        value_regno);
+                       mark_reg_unknown(state->regs, value_regno);
        } else {
                verbose("R%d invalid mem access '%s'\n",
                        regno, reg_type_str[reg->type]);
                return -EACCES;
        }
 
-       if (!err && size <= 2 && value_regno >= 0 && env->allow_ptr_leaks &&
-           state->regs[value_regno].type == UNKNOWN_VALUE) {
-               /* 1 or 2 byte load zero-extends, determine the number of
-                * zero upper bits. Not doing it fo 4 byte load, since
-                * such values cannot be added to ptr_to_packet anyway.
-                */
-               state->regs[value_regno].imm = 64 - size * 8;
+       if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
+           state->regs[value_regno].type == SCALAR_VALUE) {
+               /* b/h/w load zero-extends, mark upper bits as known 0 */
+               state->regs[value_regno].var_off = tnum_cast(
+                                       state->regs[value_regno].var_off, size);
+               /* sign bit is known zero, so we can bound the value */
+               state->regs[value_regno].min_value = 0;
+               state->regs[value_regno].max_value = min_t(u64,
+                                       state->regs[value_regno].var_off.value |
+                                       state->regs[value_regno].var_off.mask,
+                                       BPF_REGISTER_MAX_RANGE);
        }
        return err;
 }
@@ -1016,9 +1117,17 @@ static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_ins
                                BPF_SIZE(insn->code), BPF_WRITE, -1);
 }
 
+/* Does this register contain a constant zero? */
+static bool register_is_null(struct bpf_reg_state reg)
+{
+       return reg.type == SCALAR_VALUE && tnum_equals_const(reg.var_off, 0);
+}
+
 /* when register 'regno' is passed into function that will read 'access_size'
  * bytes from that pointer, make sure that it's within stack boundary
- * and all elements of stack are initialized
+ * and all elements of stack are initialized.
+ * Unlike most pointer bounds-checking functions, this one doesn't take an
+ * 'off' argument, so it has to add in reg->off itself.
  */
 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
                                int access_size, bool zero_size_allowed,
@@ -1029,9 +1138,9 @@ static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
        int off, i;
 
        if (regs[regno].type != PTR_TO_STACK) {
+               /* Allow zero-byte read from NULL, regardless of pointer type */
                if (zero_size_allowed && access_size == 0 &&
-                   regs[regno].type == CONST_IMM &&
-                   regs[regno].imm  == 0)
+                   register_is_null(regs[regno]))
                        return 0;
 
                verbose("R%d type=%s expected=%s\n", regno,
@@ -1040,7 +1149,15 @@ static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
                return -EACCES;
        }
 
-       off = regs[regno].imm;
+       /* Only allow fixed-offset stack reads */
+       if (!tnum_is_const(regs[regno].var_off)) {
+               char tn_buf[48];
+
+               tnum_strn(tn_buf, sizeof(tn_buf), regs[regno].var_off);
+               verbose("invalid variable stack read R%d var_off=%s\n",
+                       regno, tn_buf);
+       }
+       off = regs[regno].off + regs[regno].var_off.value;
        if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
            access_size <= 0) {
                verbose("invalid stack type R%d off=%d access_size=%d\n",
@@ -1071,16 +1188,14 @@ static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
                                   int access_size, bool zero_size_allowed,
                                   struct bpf_call_arg_meta *meta)
 {
-       struct bpf_reg_state *regs = env->cur_state.regs;
+       struct bpf_reg_state *regs = env->cur_state.regs, *reg = &regs[regno];
 
-       switch (regs[regno].type) {
+       switch (reg->type) {
        case PTR_TO_PACKET:
-               return check_packet_access(env, regno, 0, access_size);
+               return check_packet_access(env, regno, reg->off, access_size);
        case PTR_TO_MAP_VALUE:
-               return check_map_access(env, regno, 0, access_size);
-       case PTR_TO_MAP_VALUE_ADJ:
-               return check_map_access_adj(env, regno, 0, access_size);
-       default: /* const_imm|ptr_to_stack or invalid ptr */
+               return check_map_access(env, regno, reg->off, access_size);
+       default: /* scalar_value|ptr_to_stack or invalid ptr */
                return check_stack_boundary(env, regno, access_size,
                                            zero_size_allowed, meta);
        }
@@ -1123,11 +1238,8 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                        goto err_type;
        } else if (arg_type == ARG_CONST_SIZE ||
                   arg_type == ARG_CONST_SIZE_OR_ZERO) {
-               expected_type = CONST_IMM;
-               /* One exception. Allow UNKNOWN_VALUE registers when the
-                * boundaries are known and don't cause unsafe memory accesses
-                */
-               if (type != UNKNOWN_VALUE && type != expected_type)
+               expected_type = SCALAR_VALUE;
+               if (type != expected_type)
                        goto err_type;
        } else if (arg_type == ARG_CONST_MAP_PTR) {
                expected_type = CONST_PTR_TO_MAP;
@@ -1141,13 +1253,13 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                   arg_type == ARG_PTR_TO_UNINIT_MEM) {
                expected_type = PTR_TO_STACK;
                /* One exception here. In case function allows for NULL to be
-                * passed in as argument, it's a CONST_IMM type. Final test
+                * passed in as argument, it's a SCALAR_VALUE type. Final test
                 * happens during stack boundary checking.
                 */
-               if (type == CONST_IMM && reg->imm == 0)
+               if (register_is_null(*reg))
                        /* final test in check_stack_boundary() */;
                else if (type != PTR_TO_PACKET && type != PTR_TO_MAP_VALUE &&
-                        type != PTR_TO_MAP_VALUE_ADJ && type != expected_type)
+                        type != expected_type)
                        goto err_type;
                meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
        } else {
@@ -1173,7 +1285,7 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                        return -EACCES;
                }
                if (type == PTR_TO_PACKET)
-                       err = check_packet_access(env, regno, 0,
+                       err = check_packet_access(env, regno, reg->off,
                                                  meta->map_ptr->key_size);
                else
                        err = check_stack_boundary(env, regno,
@@ -1189,7 +1301,7 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                        return -EACCES;
                }
                if (type == PTR_TO_PACKET)
-                       err = check_packet_access(env, regno, 0,
+                       err = check_packet_access(env, regno, reg->off,
                                                  meta->map_ptr->value_size);
                else
                        err = check_stack_boundary(env, regno,
@@ -1209,10 +1321,11 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                        return -EACCES;
                }
 
-               /* If the register is UNKNOWN_VALUE, the access check happens
-                * using its boundaries. Otherwise, just use its imm
+               /* The register is SCALAR_VALUE; the access check
+                * happens using its boundaries.
                 */
-               if (type == UNKNOWN_VALUE) {
+
+               if (!tnum_is_const(reg->var_off))
                        /* For unprivileged variable accesses, disable raw
                         * mode so that the program is required to
                         * initialize all the memory that the helper could
@@ -1220,35 +1333,28 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
                         */
                        meta = NULL;
 
-                       if (reg->min_value < 0) {
-                               verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
-                                       regno);
-                               return -EACCES;
-                       }
-
-                       if (reg->min_value == 0) {
-                               err = check_helper_mem_access(env, regno - 1, 0,
-                                                             zero_size_allowed,
-                                                             meta);
-                               if (err)
-                                       return err;
-                       }
+               if (reg->min_value < 0) {
+                       verbose("R%d min value is negative, either use unsigned or 'var &= const'\n",
+                               regno);
+                       return -EACCES;
+               }
 
-                       if (reg->max_value == BPF_REGISTER_MAX_RANGE) {
-                               verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
-                                       regno);
-                               return -EACCES;
-                       }
-                       err = check_helper_mem_access(env, regno - 1,
-                                                     reg->max_value,
-                                                     zero_size_allowed, meta);
+               if (reg->min_value == 0) {
+                       err = check_helper_mem_access(env, regno - 1, 0,
+                                                     zero_size_allowed,
+                                                     meta);
                        if (err)
                                return err;
-               } else {
-                       /* register is CONST_IMM */
-                       err = check_helper_mem_access(env, regno - 1, reg->imm,
-                                                     zero_size_allowed, meta);
                }
+
+               if (reg->max_value == BPF_REGISTER_MAX_RANGE) {
+                       verbose("R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
+                               regno);
+                       return -EACCES;
+               }
+               err = check_helper_mem_access(env, regno - 1,
+                                             reg->max_value,
+                                             zero_size_allowed, meta);
        }
 
        return err;
@@ -1352,6 +1458,9 @@ static int check_raw_mode(const struct bpf_func_proto *fn)
        return count > 1 ? -EINVAL : 0;
 }
 
+/* Packet data might have moved, any old PTR_TO_PACKET[_END] are now invalid,
+ * so turn them into unknown SCALAR_VALUE.
+ */
 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
 {
        struct bpf_verifier_state *state = &env->cur_state;
@@ -1361,7 +1470,7 @@ static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
        for (i = 0; i < MAX_BPF_REG; i++)
                if (regs[i].type == PTR_TO_PACKET ||
                    regs[i].type == PTR_TO_PACKET_END)
-                       mark_reg_unknown_value(regs, i);
+                       mark_reg_unknown(regs, i);
 
        for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
                if (state->stack_slot_type[i] != STACK_SPILL)
@@ -1370,8 +1479,7 @@ static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
                if (reg->type != PTR_TO_PACKET &&
                    reg->type != PTR_TO_PACKET_END)
                        continue;
-               __mark_reg_unknown_value(state->spilled_regs,
-                                        i / BPF_REG_SIZE);
+               __mark_reg_unknown(reg);
        }
 }
 
@@ -1451,14 +1559,17 @@ static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
 
        /* update return register */
        if (fn->ret_type == RET_INTEGER) {
-               regs[BPF_REG_0].type = UNKNOWN_VALUE;
+               /* sets type to SCALAR_VALUE */
+               mark_reg_unknown(regs, BPF_REG_0);
        } else if (fn->ret_type == RET_VOID) {
                regs[BPF_REG_0].type = NOT_INIT;
        } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
                struct bpf_insn_aux_data *insn_aux;
 
                regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
-               regs[BPF_REG_0].max_value = regs[BPF_REG_0].min_value = 0;
+               /* There is no offset yet applied, variable or fixed */
+               mark_reg_known_zero(regs, BPF_REG_0);
+               regs[BPF_REG_0].off = 0;
                /* remember map_ptr, so that check_map_access()
                 * can check 'value_size' boundary of memory access
                 * to map element returned from bpf_map_lookup_elem()
@@ -1489,456 +1600,337 @@ static int check_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
        return 0;
 }
 
-static int check_packet_ptr_add(struct bpf_verifier_env *env,
-                               struct bpf_insn *insn)
+static void check_reg_overflow(struct bpf_reg_state *reg)
 {
-       struct bpf_reg_state *regs = env->cur_state.regs;
-       struct bpf_reg_state *dst_reg = &regs[insn->dst_reg];
-       struct bpf_reg_state *src_reg = &regs[insn->src_reg];
-       struct bpf_reg_state tmp_reg;
-       s32 imm;
-
-       if (BPF_SRC(insn->code) == BPF_K) {
-               /* pkt_ptr += imm */
-               imm = insn->imm;
-
-add_imm:
-               if (imm < 0) {
-                       verbose("addition of negative constant to packet pointer is not allowed\n");
-                       return -EACCES;
-               }
-               if (imm >= MAX_PACKET_OFF ||
-                   imm + dst_reg->off >= MAX_PACKET_OFF) {
-                       verbose("constant %d is too large to add to packet pointer\n",
-                               imm);
-                       return -EACCES;
-               }
-               /* a constant was added to pkt_ptr.
-                * Remember it while keeping the same 'id'
-                */
-               dst_reg->off += imm;
-       } else {
-               bool had_id;
-
-               if (src_reg->type == PTR_TO_PACKET) {
-                       /* R6=pkt(id=0,off=0,r=62) R7=imm22; r7 += r6 */
-                       tmp_reg = *dst_reg;  /* save r7 state */
-                       *dst_reg = *src_reg; /* copy pkt_ptr state r6 into r7 */
-                       src_reg = &tmp_reg;  /* pretend it's src_reg state */
-                       /* if the checks below reject it, the copy won't matter,
-                        * since we're rejecting the whole program. If all ok,
-                        * then imm22 state will be added to r7
-                        * and r7 will be pkt(id=0,off=22,r=62) while
-                        * r6 will stay as pkt(id=0,off=0,r=62)
-                        */
-               }
-
-               if (src_reg->type == CONST_IMM) {
-                       /* pkt_ptr += reg where reg is known constant */
-                       imm = src_reg->imm;
-                       goto add_imm;
-               }
-               /* disallow pkt_ptr += reg
-                * if reg is not uknown_value with guaranteed zero upper bits
-                * otherwise pkt_ptr may overflow and addition will become
-                * subtraction which is not allowed
-                */
-               if (src_reg->type != UNKNOWN_VALUE) {
-                       verbose("cannot add '%s' to ptr_to_packet\n",
-                               reg_type_str[src_reg->type]);
-                       return -EACCES;
-               }
-               if (src_reg->imm < 48) {
-                       verbose("cannot add integer value with %lld upper zero bits to ptr_to_packet\n",
-                               src_reg->imm);
-                       return -EACCES;
-               }
-
-               had_id = (dst_reg->id != 0);
+       if (reg->max_value > BPF_REGISTER_MAX_RANGE)
+               reg->max_value = BPF_REGISTER_MAX_RANGE;
+       if (reg->min_value < BPF_REGISTER_MIN_RANGE ||
+           reg->min_value > BPF_REGISTER_MAX_RANGE)
+               reg->min_value = BPF_REGISTER_MIN_RANGE;
+}
 
-               /* dst_reg stays as pkt_ptr type and since some positive
-                * integer value was added to the pointer, increment its 'id'
-                */
-               dst_reg->id = ++env->id_gen;
-
-               /* something was added to pkt_ptr, set range to zero */
-               dst_reg->aux_off += dst_reg->off;
-               dst_reg->off = 0;
-               dst_reg->range = 0;
-               if (had_id)
-                       dst_reg->aux_off_align = min(dst_reg->aux_off_align,
-                                                    src_reg->min_align);
-               else
-                       dst_reg->aux_off_align = src_reg->min_align;
+static void coerce_reg_to_32(struct bpf_reg_state *reg)
+{
+       /* 32-bit values can't be negative as an s64 */
+       if (reg->min_value < 0)
+               reg->min_value = 0;
+       /* clear high 32 bits */
+       reg->var_off = tnum_cast(reg->var_off, 4);
+       /* Did value become known?  Then update bounds */
+       if (tnum_is_const(reg->var_off)) {
+               if ((s64)reg->var_off.value > BPF_REGISTER_MIN_RANGE)
+                       reg->min_value = reg->var_off.value;
+               if (reg->var_off.value < BPF_REGISTER_MAX_RANGE)
+                       reg->max_value = reg->var_off.value;
        }
-       return 0;
 }
 
-static int evaluate_reg_alu(struct bpf_verifier_env *env, struct bpf_insn *insn)
+/* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
+ * Caller must check_reg_overflow all argument regs beforehand.
+ * Caller should also handle BPF_MOV case separately.
+ * If we return -EACCES, caller may want to try again treating pointer as a
+ * scalar.  So we only emit a diagnostic if !env->allow_ptr_leaks.
+ */
+static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
+                                  struct bpf_insn *insn,
+                                  const struct bpf_reg_state *ptr_reg,
+                                  const struct bpf_reg_state *off_reg)
 {
-       struct bpf_reg_state *regs = env->cur_state.regs;
-       struct bpf_reg_state *dst_reg = &regs[insn->dst_reg];
+       struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
+       bool known = tnum_is_const(off_reg->var_off);
+       s64 min_val = off_reg->min_value;
+       u64 max_val = off_reg->max_value;
        u8 opcode = BPF_OP(insn->code);
-       s64 imm_log2;
+       u32 dst = insn->dst_reg;
 
-       /* for type == UNKNOWN_VALUE:
-        * imm > 0 -> number of zero upper bits
-        * imm == 0 -> don't track which is the same as all bits can be non-zero
-        */
+       dst_reg = &regs[dst];
 
-       if (BPF_SRC(insn->code) == BPF_X) {
-               struct bpf_reg_state *src_reg = &regs[insn->src_reg];
-
-               if (src_reg->type == UNKNOWN_VALUE && src_reg->imm > 0 &&
-                   dst_reg->imm && opcode == BPF_ADD) {
-                       /* dreg += sreg
-                        * where both have zero upper bits. Adding them
-                        * can only result making one more bit non-zero
-                        * in the larger value.
-                        * Ex. 0xffff (imm=48) + 1 (imm=63) = 0x10000 (imm=47)
-                        *     0xffff (imm=48) + 0xffff = 0x1fffe (imm=47)
-                        */
-                       dst_reg->imm = min(dst_reg->imm, src_reg->imm);
-                       dst_reg->imm--;
-                       return 0;
-               }
-               if (src_reg->type == CONST_IMM && src_reg->imm > 0 &&
-                   dst_reg->imm && opcode == BPF_ADD) {
-                       /* dreg += sreg
-                        * where dreg has zero upper bits and sreg is const.
-                        * Adding them can only result making one more bit
-                        * non-zero in the larger value.
-                        */
-                       imm_log2 = __ilog2_u64((long long)src_reg->imm);
-                       dst_reg->imm = min(dst_reg->imm, 63 - imm_log2);
-                       dst_reg->imm--;
-                       return 0;
-               }
-               /* all other cases non supported yet, just mark dst_reg */
-               dst_reg->imm = 0;
-               return 0;
+       if (WARN_ON_ONCE(known && (min_val != max_val))) {
+               print_verifier_state(&env->cur_state);
+               verbose("verifier internal error\n");
+               return -EINVAL;
+       }
+
+       if (BPF_CLASS(insn->code) != BPF_ALU64) {
+               /* 32-bit ALU ops on pointers produce (meaningless) scalars */
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d 32-bit pointer arithmetic prohibited\n",
+                               dst);
+               return -EACCES;
        }
 
-       /* sign extend 32-bit imm into 64-bit to make sure that
-        * negative values occupy bit 63. Note ilog2() would have
-        * been incorrect, since sizeof(insn->imm) == 4
+       if (ptr_reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d pointer arithmetic on PTR_TO_MAP_VALUE_OR_NULL prohibited, null-check it first\n",
+                               dst);
+               return -EACCES;
+       }
+       if (ptr_reg->type == CONST_PTR_TO_MAP) {
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d pointer arithmetic on CONST_PTR_TO_MAP prohibited\n",
+                               dst);
+               return -EACCES;
+       }
+       if (ptr_reg->type == PTR_TO_PACKET_END) {
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d pointer arithmetic on PTR_TO_PACKET_END prohibited\n",
+                               dst);
+               return -EACCES;
+       }
+
+       /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
+        * The id may be overwritten later if we create a new variable offset.
         */
-       imm_log2 = __ilog2_u64((long long)insn->imm);
+       dst_reg->type = ptr_reg->type;
+       dst_reg->id = ptr_reg->id;
 
-       if (dst_reg->imm && opcode == BPF_LSH) {
-               /* reg <<= imm
-                * if reg was a result of 2 byte load, then its imm == 48
-                * which means that upper 48 bits are zero and shifting this reg
-                * left by 4 would mean that upper 44 bits are still zero
+       switch (opcode) {
+       case BPF_ADD:
+               /* We can take a fixed offset as long as it doesn't overflow
+                * the s32 'off' field
                 */
-               dst_reg->imm -= insn->imm;
-       } else if (dst_reg->imm && opcode == BPF_MUL) {
-               /* reg *= imm
-                * if multiplying by 14 subtract 4
-                * This is conservative calculation of upper zero bits.
-                * It's not trying to special case insn->imm == 1 or 0 cases
+               if (known && (ptr_reg->off + min_val ==
+                             (s64)(s32)(ptr_reg->off + min_val))) {
+                       /* pointer += K.  Accumulate it into fixed offset */
+                       dst_reg->min_value = ptr_reg->min_value;
+                       dst_reg->max_value = ptr_reg->max_value;
+                       dst_reg->var_off = ptr_reg->var_off;
+                       dst_reg->off = ptr_reg->off + min_val;
+                       dst_reg->range = ptr_reg->range;
+                       break;
+               }
+               if (max_val == BPF_REGISTER_MAX_RANGE) {
+                       if (!env->allow_ptr_leaks)
+                               verbose("R%d tried to add unbounded value to pointer\n",
+                                       dst);
+                       return -EACCES;
+               }
+               /* A new variable offset is created.  Note that off_reg->off
+                * == 0, since it's a scalar.
+                * dst_reg gets the pointer type and since some positive
+                * integer value was added to the pointer, give it a new 'id'
+                * if it's a PTR_TO_PACKET.
+                * this creates a new 'base' pointer, off_reg (variable) gets
+                * added into the variable offset, and we copy the fixed offset
+                * from ptr_reg.
                 */
-               dst_reg->imm -= imm_log2 + 1;
-       } else if (opcode == BPF_AND) {
-               /* reg &= imm */
-               dst_reg->imm = 63 - imm_log2;
-       } else if (dst_reg->imm && opcode == BPF_ADD) {
-               /* reg += imm */
-               dst_reg->imm = min(dst_reg->imm, 63 - imm_log2);
-               dst_reg->imm--;
-       } else if (opcode == BPF_RSH) {
-               /* reg >>= imm
-                * which means that after right shift, upper bits will be zero
-                * note that verifier already checked that
-                * 0 <= imm < 64 for shift insn
+               if (min_val <= BPF_REGISTER_MIN_RANGE)
+                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
+               if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
+                       dst_reg->min_value += min_val;
+               if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
+                       dst_reg->max_value += max_val;
+               dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
+               dst_reg->off = ptr_reg->off;
+               if (ptr_reg->type == PTR_TO_PACKET) {
+                       dst_reg->id = ++env->id_gen;
+                       /* something was added to pkt_ptr, set range to zero */
+                       dst_reg->range = 0;
+               }
+               break;
+       case BPF_SUB:
+               if (dst_reg == off_reg) {
+                       /* scalar -= pointer.  Creates an unknown scalar */
+                       if (!env->allow_ptr_leaks)
+                               verbose("R%d tried to subtract pointer from scalar\n",
+                                       dst);
+                       return -EACCES;
+               }
+               /* We don't allow subtraction from FP, because (according to
+                * test_verifier.c test "invalid fp arithmetic", JITs might not
+                * be able to deal with it.
                 */
-               dst_reg->imm += insn->imm;
-               if (unlikely(dst_reg->imm > 64))
-                       /* some dumb code did:
-                        * r2 = *(u32 *)mem;
-                        * r2 >>= 32;
-                        * and all bits are zero now */
-                       dst_reg->imm = 64;
-       } else {
-               /* all other alu ops, means that we don't know what will
-                * happen to the value, mark it with unknown number of zero bits
+               if (ptr_reg->type == PTR_TO_STACK) {
+                       if (!env->allow_ptr_leaks)
+                               verbose("R%d subtraction from stack pointer prohibited\n",
+                                       dst);
+                       return -EACCES;
+               }
+               if (known && (ptr_reg->off - min_val ==
+                             (s64)(s32)(ptr_reg->off - min_val))) {
+                       /* pointer -= K.  Subtract it from fixed offset */
+                       dst_reg->min_value = ptr_reg->min_value;
+                       dst_reg->max_value = ptr_reg->max_value;
+                       dst_reg->var_off = ptr_reg->var_off;
+                       dst_reg->id = ptr_reg->id;
+                       dst_reg->off = ptr_reg->off - min_val;
+                       dst_reg->range = ptr_reg->range;
+                       break;
+               }
+               /* Subtracting a negative value will just confuse everything.
+                * This can happen if off_reg is an immediate.
                 */
-               dst_reg->imm = 0;
-       }
-
-       if (dst_reg->imm < 0) {
-               /* all 64 bits of the register can contain non-zero bits
-                * and such value cannot be added to ptr_to_packet, since it
-                * may overflow, mark it as unknown to avoid further eval
+               if ((s64)max_val < 0) {
+                       if (!env->allow_ptr_leaks)
+                               verbose("R%d tried to subtract negative max_val %lld from pointer\n",
+                                       dst, (s64)max_val);
+                       return -EACCES;
+               }
+               /* A new variable offset is created.  If the subtrahend is known
+                * nonnegative, then any reg->range we had before is still good.
                 */
-               dst_reg->imm = 0;
-       }
-       return 0;
-}
-
-static int evaluate_reg_imm_alu_unknown(struct bpf_verifier_env *env,
-                                       struct bpf_insn *insn)
-{
-       struct bpf_reg_state *regs = env->cur_state.regs;
-       struct bpf_reg_state *dst_reg = &regs[insn->dst_reg];
-       struct bpf_reg_state *src_reg = &regs[insn->src_reg];
-       u8 opcode = BPF_OP(insn->code);
-       s64 imm_log2 = __ilog2_u64((long long)dst_reg->imm);
-
-       /* BPF_X code with src_reg->type UNKNOWN_VALUE here. */
-       if (src_reg->imm > 0 && dst_reg->imm) {
-               switch (opcode) {
-               case BPF_ADD:
-                       /* dreg += sreg
-                        * where both have zero upper bits. Adding them
-                        * can only result making one more bit non-zero
-                        * in the larger value.
-                        * Ex. 0xffff (imm=48) + 1 (imm=63) = 0x10000 (imm=47)
-                        *     0xffff (imm=48) + 0xffff = 0x1fffe (imm=47)
-                        */
-                       dst_reg->imm = min(src_reg->imm, 63 - imm_log2);
-                       dst_reg->imm--;
-                       break;
-               case BPF_AND:
-                       /* dreg &= sreg
-                        * AND can not extend zero bits only shrink
-                        * Ex.  0x00..00ffffff
-                        *    & 0x0f..ffffffff
-                        *     ----------------
-                        *      0x00..00ffffff
-                        */
-                       dst_reg->imm = max(src_reg->imm, 63 - imm_log2);
-                       break;
-               case BPF_OR:
-                       /* dreg |= sreg
-                        * OR can only extend zero bits
-                        * Ex.  0x00..00ffffff
-                        *    | 0x0f..ffffffff
-                        *     ----------------
-                        *      0x0f..00ffffff
-                        */
-                       dst_reg->imm = min(src_reg->imm, 63 - imm_log2);
-                       break;
-               case BPF_SUB:
-               case BPF_MUL:
-               case BPF_RSH:
-               case BPF_LSH:
-                       /* These may be flushed out later */
-               default:
-                       mark_reg_unknown_value(regs, insn->dst_reg);
+               if (max_val >= BPF_REGISTER_MAX_RANGE)
+                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
+               if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
+                       dst_reg->min_value -= max_val;
+               if (min_val <= BPF_REGISTER_MIN_RANGE)
+                       dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
+               if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
+                       dst_reg->max_value -= min_val;
+               dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
+               dst_reg->off = ptr_reg->off;
+               if (ptr_reg->type == PTR_TO_PACKET) {
+                       dst_reg->id = ++env->id_gen;
+                       /* something was added to pkt_ptr, set range to zero */
+                       if (min_val < 0)
+                               dst_reg->range = 0;
                }
-       } else {
-               mark_reg_unknown_value(regs, insn->dst_reg);
+               break;
+       case BPF_AND:
+       case BPF_OR:
+       case BPF_XOR:
+               /* bitwise ops on pointers are troublesome, prohibit for now.
+                * (However, in principle we could allow some cases, e.g.
+                * ptr &= ~3 which would reduce min_value by 3.)
+                */
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d bitwise operator %s on pointer prohibited\n",
+                               dst, bpf_alu_string[opcode >> 4]);
+               return -EACCES;
+       default:
+               /* other operators (e.g. MUL,LSH) produce non-pointer results */
+               if (!env->allow_ptr_leaks)
+                       verbose("R%d pointer arithmetic with %s operator prohibited\n",
+                               dst, bpf_alu_string[opcode >> 4]);
+               return -EACCES;
        }
 
-       dst_reg->type = UNKNOWN_VALUE;
+       check_reg_overflow(dst_reg);
        return 0;
 }
 
-static int evaluate_reg_imm_alu(struct bpf_verifier_env *env,
-                               struct bpf_insn *insn)
+static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
+                                     struct bpf_insn *insn,
+                                     struct bpf_reg_state *dst_reg,
+                                     struct bpf_reg_state src_reg)
 {
        struct bpf_reg_state *regs = env->cur_state.regs;
-       struct bpf_reg_state *dst_reg = &regs[insn->dst_reg];
-       struct bpf_reg_state *src_reg = &regs[insn->src_reg];
-       u8 opcode = BPF_OP(insn->code);
-       u64 dst_imm = dst_reg->imm;
-
-       if (BPF_SRC(insn->code) == BPF_X && src_reg->type == UNKNOWN_VALUE)
-               return evaluate_reg_imm_alu_unknown(env, insn);
-
-       /* dst_reg->type == CONST_IMM here. Simulate execution of insns
-        * containing ALU ops. Don't care about overflow or negative
-        * values, just add/sub/... them; registers are in u64.
-        */
-       if (opcode == BPF_ADD && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm += insn->imm;
-       } else if (opcode == BPF_ADD && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm += src_reg->imm;
-       } else if (opcode == BPF_SUB && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm -= insn->imm;
-       } else if (opcode == BPF_SUB && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm -= src_reg->imm;
-       } else if (opcode == BPF_MUL && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm *= insn->imm;
-       } else if (opcode == BPF_MUL && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm *= src_reg->imm;
-       } else if (opcode == BPF_OR && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm |= insn->imm;
-       } else if (opcode == BPF_OR && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm |= src_reg->imm;
-       } else if (opcode == BPF_AND && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm &= insn->imm;
-       } else if (opcode == BPF_AND && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm &= src_reg->imm;
-       } else if (opcode == BPF_RSH && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm >>= insn->imm;
-       } else if (opcode == BPF_RSH && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm >>= src_reg->imm;
-       } else if (opcode == BPF_LSH && BPF_SRC(insn->code) == BPF_K) {
-               dst_imm <<= insn->imm;
-       } else if (opcode == BPF_LSH && BPF_SRC(insn->code) == BPF_X &&
-                  src_reg->type == CONST_IMM) {
-               dst_imm <<= src_reg->imm;
-       } else {
-               mark_reg_unknown_value(regs, insn->dst_reg);
-               goto out;
-       }
-
-       dst_reg->imm = dst_imm;
-out:
-       return 0;
-}
-
-static void check_reg_overflow(struct bpf_reg_state *reg)
-{
-       if (reg->max_value > BPF_REGISTER_MAX_RANGE)
-               reg->max_value = BPF_REGISTER_MAX_RANGE;
-       if (reg->min_value < BPF_REGISTER_MIN_RANGE ||
-           reg->min_value > BPF_REGISTER_MAX_RANGE)
-               reg->min_value = BPF_REGISTER_MIN_RANGE;
-}
-
-static u32 calc_align(u32 imm)
-{
-       if (!imm)
-               return 1U << 31;
-       return imm - ((imm - 1) & imm);
-}
-
-static void adjust_reg_min_max_vals(struct bpf_verifier_env *env,
-                                   struct bpf_insn *insn)
-{
-       struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
        s64 min_val = BPF_REGISTER_MIN_RANGE;
        u64 max_val = BPF_REGISTER_MAX_RANGE;
        u8 opcode = BPF_OP(insn->code);
-       u32 dst_align, src_align;
+       bool src_known, dst_known;
 
-       dst_reg = &regs[insn->dst_reg];
-       src_align = 0;
-       if (BPF_SRC(insn->code) == BPF_X) {
-               check_reg_overflow(&regs[insn->src_reg]);
-               min_val = regs[insn->src_reg].min_value;
-               max_val = regs[insn->src_reg].max_value;
-
-               /* If the source register is a random pointer then the
-                * min_value/max_value values represent the range of the known
-                * accesses into that value, not the actual min/max value of the
-                * register itself.  In this case we have to reset the reg range
-                * values so we know it is not safe to look at.
-                */
-               if (regs[insn->src_reg].type != CONST_IMM &&
-                   regs[insn->src_reg].type != UNKNOWN_VALUE) {
-                       min_val = BPF_REGISTER_MIN_RANGE;
-                       max_val = BPF_REGISTER_MAX_RANGE;
-                       src_align = 0;
-               } else {
-                       src_align = regs[insn->src_reg].min_align;
-               }
-       } else if (insn->imm < BPF_REGISTER_MAX_RANGE &&
-                  (s64)insn->imm > BPF_REGISTER_MIN_RANGE) {
-               min_val = max_val = insn->imm;
-               src_align = calc_align(insn->imm);
-       }
-
-       dst_align = dst_reg->min_align;
-
-       /* We don't know anything about what was done to this register, mark it
-        * as unknown. Also, if both derived bounds came from signed/unsigned
-        * mixed compares and one side is unbounded, we cannot really do anything
-        * with them as boundaries cannot be trusted. Thus, arithmetic of two
-        * regs of such kind will get invalidated bounds on the dst side.
-        */
-       if ((min_val == BPF_REGISTER_MIN_RANGE &&
-            max_val == BPF_REGISTER_MAX_RANGE) ||
-           (BPF_SRC(insn->code) == BPF_X &&
-            ((min_val != BPF_REGISTER_MIN_RANGE &&
-              max_val == BPF_REGISTER_MAX_RANGE) ||
-             (min_val == BPF_REGISTER_MIN_RANGE &&
-              max_val != BPF_REGISTER_MAX_RANGE) ||
-             (dst_reg->min_value != BPF_REGISTER_MIN_RANGE &&
-              dst_reg->max_value == BPF_REGISTER_MAX_RANGE) ||
-             (dst_reg->min_value == BPF_REGISTER_MIN_RANGE &&
-              dst_reg->max_value != BPF_REGISTER_MAX_RANGE)) &&
-            regs[insn->dst_reg].value_from_signed !=
-            regs[insn->src_reg].value_from_signed)) {
-               reset_reg_range_values(regs, insn->dst_reg);
-               return;
-       }
-
-       /* If one of our values was at the end of our ranges then we can't just
-        * do our normal operations to the register, we need to set the values
-        * to the min/max since they are undefined.
-        */
-       if (opcode != BPF_SUB) {
-               if (min_val == BPF_REGISTER_MIN_RANGE)
-                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
-               if (max_val == BPF_REGISTER_MAX_RANGE)
-                       dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
+       if (BPF_CLASS(insn->code) != BPF_ALU64) {
+               /* 32-bit ALU ops are (32,32)->64 */
+               coerce_reg_to_32(dst_reg);
+               coerce_reg_to_32(&src_reg);
        }
+       min_val = src_reg.min_value;
+       max_val = src_reg.max_value;
+       src_known = tnum_is_const(src_reg.var_off);
+       dst_known = tnum_is_const(dst_reg->var_off);
 
        switch (opcode) {
        case BPF_ADD:
+               if (min_val == BPF_REGISTER_MIN_RANGE)
+                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
                if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
                        dst_reg->min_value += min_val;
+               /* if max_val is MAX_RANGE, this will saturate dst->max */
                if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
                        dst_reg->max_value += max_val;
-               dst_reg->min_align = min(src_align, dst_align);
+               dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
                break;
        case BPF_SUB:
-               /* If one of our values was at the end of our ranges, then the
-                * _opposite_ value in the dst_reg goes to the end of our range.
-                */
-               if (min_val == BPF_REGISTER_MIN_RANGE)
-                       dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
                if (max_val == BPF_REGISTER_MAX_RANGE)
                        dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
                if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
                        dst_reg->min_value -= max_val;
+               if (min_val == BPF_REGISTER_MIN_RANGE)
+                       dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
                if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
                        dst_reg->max_value -= min_val;
-               dst_reg->min_align = min(src_align, dst_align);
+               dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
                break;
        case BPF_MUL:
-               if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
-                       dst_reg->min_value *= min_val;
+               if (min_val < 0 || dst_reg->min_value < 0) {
+                       /* Ain't nobody got time to multiply that sign */
+                       __mark_reg_unknown(dst_reg);
+                       break;
+               }
+               dst_reg->min_value *= min_val;
+               /* if max_val is MAX_RANGE, this will saturate dst->max.
+                * We know MAX_RANGE ** 2 won't overflow a u64, because
+                * MAX_RANGE itself fits in a u32.
+                */
+               BUILD_BUG_ON(BPF_REGISTER_MAX_RANGE > (u32)-1);
                if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
                        dst_reg->max_value *= max_val;
-               dst_reg->min_align = max(src_align, dst_align);
+               dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
                break;
        case BPF_AND:
-               /* Disallow AND'ing of negative numbers, ain't nobody got time
-                * for that.  Otherwise the minimum is 0 and the max is the max
-                * value we could AND against.
+               if (src_known && dst_known) {
+                       u64 value = dst_reg->var_off.value & src_reg.var_off.value;
+
+                       dst_reg->var_off = tnum_const(value);
+                       dst_reg->min_value = dst_reg->max_value = min_t(u64,
+                                       value, BPF_REGISTER_MAX_RANGE);
+                       break;
+               }
+               /* Lose min_value when AND'ing negative numbers, ain't nobody
+                * got time for that.  Otherwise we get our minimum from the
+                * var_off, since that's inherently bitwise.
+                * Our maximum is the minimum of the operands' maxima.
                 */
-               if (min_val < 0)
+               dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
+               if (min_val < 0 && dst_reg->min_value < 0)
                        dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
                else
-                       dst_reg->min_value = 0;
-               dst_reg->max_value = max_val;
-               dst_reg->min_align = max(src_align, dst_align);
+                       dst_reg->min_value = dst_reg->var_off.value;
+               dst_reg->max_value = min(dst_reg->max_value, max_val);
+               break;
+       case BPF_OR:
+               if (src_known && dst_known) {
+                       u64 value = dst_reg->var_off.value | src_reg.var_off.value;
+
+                       dst_reg->var_off = tnum_const(value);
+                       dst_reg->min_value = dst_reg->max_value = min_t(u64,
+                                       value, BPF_REGISTER_MAX_RANGE);
+                       break;
+               }
+               /* Lose ranges when OR'ing negative numbers, ain't nobody got
+                * time for that.  Otherwise we get our maximum from the var_off,
+                * and our minimum is the maximum of the operands' minima.
+                */
+               dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
+               if (min_val < 0 || dst_reg->min_value < 0) {
+                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
+                       dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
+               } else {
+                       dst_reg->min_value = max(dst_reg->min_value, min_val);
+                       dst_reg->max_value = dst_reg->var_off.value | dst_reg->var_off.mask;
+               }
                break;
        case BPF_LSH:
+               if (min_val < 0) {
+                       /* LSH by a negative number is undefined */
+                       mark_reg_unknown(regs, insn->dst_reg);
+                       break;
+               }
                /* Gotta have special overflow logic here, if we're shifting
                 * more than MAX_RANGE then just assume we have an invalid
                 * range.
                 */
                if (min_val > ilog2(BPF_REGISTER_MAX_RANGE)) {
                        dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
-                       dst_reg->min_align = 1;
+                       dst_reg->var_off = tnum_unknown;
                } else {
                        if (dst_reg->min_value != BPF_REGISTER_MIN_RANGE)
                                dst_reg->min_value <<= min_val;
-                       if (!dst_reg->min_align)
-                               dst_reg->min_align = 1;
-                       dst_reg->min_align <<= min_val;
+                       if (src_known)
+                               dst_reg->var_off = tnum_lshift(dst_reg->var_off, min_val);
+                       else
+                               dst_reg->var_off = tnum_lshift(tnum_unknown, min_val);
                }
                if (max_val > ilog2(BPF_REGISTER_MAX_RANGE))
                        dst_reg->max_value = BPF_REGISTER_MAX_RANGE;
@@ -1946,37 +1938,139 @@ static void adjust_reg_min_max_vals(struct bpf_verifier_env *env,
                        dst_reg->max_value <<= max_val;
                break;
        case BPF_RSH:
-               /* RSH by a negative number is undefined, and the BPF_RSH is an
-                * unsigned shift, so make the appropriate casts.
-                */
-               if (min_val < 0 || dst_reg->min_value < 0) {
-                       dst_reg->min_value = BPF_REGISTER_MIN_RANGE;
+               if (min_val < 0) {
+                       /* RSH by a negative number is undefined */
+                       mark_reg_unknown(regs, insn->dst_reg);
+                       break;
+               }
+               /* BPF_RSH is an unsigned shift, so make the appropriate casts */
+               if (dst_reg->min_value < 0) {
+                       if (min_val)
+                               /* Sign bit will be cleared */
+                               dst_reg->min_value = 0;
                } else {
                        dst_reg->min_value =
                                (u64)(dst_reg->min_value) >> min_val;
                }
-               if (min_val < 0) {
-                       dst_reg->min_align = 1;
-               } else {
-                       dst_reg->min_align >>= (u64) min_val;
-                       if (!dst_reg->min_align)
-                               dst_reg->min_align = 1;
-               }
-               if (dst_reg->max_value != BPF_REGISTER_MAX_RANGE)
-                       dst_reg->max_value >>= max_val;
+               if (src_known)
+                       dst_reg->var_off = tnum_rshift(dst_reg->var_off, min_val);
+               else
+                       dst_reg->var_off = tnum_rshift(tnum_unknown, min_val);
+               if (dst_reg->max_value == BPF_REGISTER_MAX_RANGE)
+                       dst_reg->max_value = ~0;
+               dst_reg->max_value >>= max_val;
                break;
        default:
-               reset_reg_range_values(regs, insn->dst_reg);
+               mark_reg_unknown(regs, insn->dst_reg);
                break;
        }
 
        check_reg_overflow(dst_reg);
+       return 0;
+}
+
+/* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
+ * and var_off.
+ */
+static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
+                                  struct bpf_insn *insn)
+{
+       struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg, *src_reg;
+       struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
+       u8 opcode = BPF_OP(insn->code);
+       int rc;
+
+       dst_reg = &regs[insn->dst_reg];
+       check_reg_overflow(dst_reg);
+       src_reg = NULL;
+       if (dst_reg->type != SCALAR_VALUE)
+               ptr_reg = dst_reg;
+       if (BPF_SRC(insn->code) == BPF_X) {
+               src_reg = &regs[insn->src_reg];
+               check_reg_overflow(src_reg);
+
+               if (src_reg->type != SCALAR_VALUE) {
+                       if (dst_reg->type != SCALAR_VALUE) {
+                               /* Combining two pointers by any ALU op yields
+                                * an arbitrary scalar.
+                                */
+                               if (!env->allow_ptr_leaks) {
+                                       verbose("R%d pointer %s pointer prohibited\n",
+                                               insn->dst_reg,
+                                               bpf_alu_string[opcode >> 4]);
+                                       return -EACCES;
+                               }
+                               mark_reg_unknown(regs, insn->dst_reg);
+                               return 0;
+                       } else {
+                               /* scalar += pointer
+                                * This is legal, but we have to reverse our
+                                * src/dest handling in computing the range
+                                */
+                               rc = adjust_ptr_min_max_vals(env, insn,
+                                                            src_reg, dst_reg);
+                               if (rc == -EACCES && env->allow_ptr_leaks) {
+                                       /* scalar += unknown scalar */
+                                       __mark_reg_unknown(&off_reg);
+                                       return adjust_scalar_min_max_vals(
+                                                       env, insn,
+                                                       dst_reg, off_reg);
+                               }
+                               return rc;
+                       }
+               } else if (ptr_reg) {
+                       /* pointer += scalar */
+                       rc = adjust_ptr_min_max_vals(env, insn,
+                                                    dst_reg, src_reg);
+                       if (rc == -EACCES && env->allow_ptr_leaks) {
+                               /* unknown scalar += scalar */
+                               __mark_reg_unknown(dst_reg);
+                               return adjust_scalar_min_max_vals(
+                                               env, insn, dst_reg, *src_reg);
+                       }
+                       return rc;
+               }
+       } else {
+               /* Pretend the src is a reg with a known value, since we only
+                * need to be able to read from this state.
+                */
+               off_reg.type = SCALAR_VALUE;
+               off_reg.var_off = tnum_const(insn->imm);
+               off_reg.min_value = insn->imm;
+               off_reg.max_value = insn->imm;
+               src_reg = &off_reg;
+               check_reg_overflow(src_reg);
+               if (ptr_reg) { /* pointer += K */
+                       rc = adjust_ptr_min_max_vals(env, insn,
+                                                    ptr_reg, src_reg);
+                       if (rc == -EACCES && env->allow_ptr_leaks) {
+                               /* unknown scalar += K */
+                               __mark_reg_unknown(dst_reg);
+                               return adjust_scalar_min_max_vals(
+                                               env, insn, dst_reg, off_reg);
+                       }
+                       return rc;
+               }
+       }
+
+       /* Got here implies adding two SCALAR_VALUEs */
+       if (WARN_ON_ONCE(ptr_reg)) {
+               print_verifier_state(&env->cur_state);
+               verbose("verifier internal error: unexpected ptr_reg\n");
+               return -EINVAL;
+       }
+       if (WARN_ON(!src_reg)) {
+               print_verifier_state(&env->cur_state);
+               verbose("verifier internal error: no src_reg\n");
+               return -EINVAL;
+       }
+       return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
 }
 
 /* check validity of 32-bit and 64-bit arithmetic operations */
 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
 {
-       struct bpf_reg_state *regs = env->cur_state.regs, *dst_reg;
+       struct bpf_reg_state *regs = env->cur_state.regs;
        u8 opcode = BPF_OP(insn->code);
        int err;
 
@@ -2036,11 +2130,6 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
                if (err)
                        return err;
 
-               /* we are setting our register to something new, we need to
-                * reset its range values.
-                */
-               reset_reg_range_values(regs, insn->dst_reg);
-
                if (BPF_SRC(insn->code) == BPF_X) {
                        if (BPF_CLASS(insn->code) == BPF_ALU64) {
                                /* case: R1 = R2
@@ -2048,24 +2137,29 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
                                 */
                                regs[insn->dst_reg] = regs[insn->src_reg];
                        } else {
+                               /* R1 = (u32) R2 */
                                if (is_pointer_value(env, insn->src_reg)) {
                                        verbose("R%d partial copy of pointer\n",
                                                insn->src_reg);
                                        return -EACCES;
                                }
-                               mark_reg_unknown_value(regs, insn->dst_reg);
+                               mark_reg_unknown(regs, insn->dst_reg);
+                               /* high 32 bits are known zero.  But this is
+                                * still out of range for max_value, so leave
+                                * that.
+                                */
+                               regs[insn->dst_reg].var_off = tnum_cast(
+                                               regs[insn->dst_reg].var_off, 4);
                        }
                } else {
                        /* case: R = imm
                         * remember the value we stored into this reg
                         */
-                       regs[insn->dst_reg].type = CONST_IMM;
-                       regs[insn->dst_reg].imm = insn->imm;
-                       regs[insn->dst_reg].id = 0;
+                       regs[insn->dst_reg].type = SCALAR_VALUE;
+                       regs[insn->dst_reg].var_off = tnum_const(insn->imm);
                        regs[insn->dst_reg].max_value = insn->imm;
                        regs[insn->dst_reg].min_value = insn->imm;
-                       regs[insn->dst_reg].min_align = calc_align(insn->imm);
-                       regs[insn->dst_reg].value_from_signed = false;
+                       regs[insn->dst_reg].id = 0;
                }
 
        } else if (opcode > BPF_END) {
@@ -2116,68 +2210,7 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
                if (err)
                        return err;
 
-               dst_reg = &regs[insn->dst_reg];
-
-               /* first we want to adjust our ranges. */
-               adjust_reg_min_max_vals(env, insn);
-
-               /* pattern match 'bpf_add Rx, imm' instruction */
-               if (opcode == BPF_ADD && BPF_CLASS(insn->code) == BPF_ALU64 &&
-                   dst_reg->type == FRAME_PTR && BPF_SRC(insn->code) == BPF_K) {
-                       dst_reg->type = PTR_TO_STACK;
-                       dst_reg->imm = insn->imm;
-                       return 0;
-               } else if (opcode == BPF_ADD &&
-                          BPF_CLASS(insn->code) == BPF_ALU64 &&
-                          dst_reg->type == PTR_TO_STACK &&
-                          ((BPF_SRC(insn->code) == BPF_X &&
-                            regs[insn->src_reg].type == CONST_IMM) ||
-                           BPF_SRC(insn->code) == BPF_K)) {
-                       if (BPF_SRC(insn->code) == BPF_X)
-                               dst_reg->imm += regs[insn->src_reg].imm;
-                       else
-                               dst_reg->imm += insn->imm;
-                       return 0;
-               } else if (opcode == BPF_ADD &&
-                          BPF_CLASS(insn->code) == BPF_ALU64 &&
-                          (dst_reg->type == PTR_TO_PACKET ||
-                           (BPF_SRC(insn->code) == BPF_X &&
-                            regs[insn->src_reg].type == PTR_TO_PACKET))) {
-                       /* ptr_to_packet += K|X */
-                       return check_packet_ptr_add(env, insn);
-               } else if (BPF_CLASS(insn->code) == BPF_ALU64 &&
-                          dst_reg->type == UNKNOWN_VALUE &&
-                          env->allow_ptr_leaks) {
-                       /* unknown += K|X */
-                       return evaluate_reg_alu(env, insn);
-               } else if (BPF_CLASS(insn->code) == BPF_ALU64 &&
-                          dst_reg->type == CONST_IMM &&
-                          env->allow_ptr_leaks) {
-                       /* reg_imm += K|X */
-                       return evaluate_reg_imm_alu(env, insn);
-               } else if (is_pointer_value(env, insn->dst_reg)) {
-                       verbose("R%d pointer arithmetic prohibited\n",
-                               insn->dst_reg);
-                       return -EACCES;
-               } else if (BPF_SRC(insn->code) == BPF_X &&
-                          is_pointer_value(env, insn->src_reg)) {
-                       verbose("R%d pointer arithmetic prohibited\n",
-                               insn->src_reg);
-                       return -EACCES;
-               }
-
-               /* If we did pointer math on a map value then just set it to our
-                * PTR_TO_MAP_VALUE_ADJ type so we can deal with any stores or
-                * loads to this register appropriately, otherwise just mark the
-                * register as unknown.
-                */
-               if (env->allow_ptr_leaks &&
-                   BPF_CLASS(insn->code) == BPF_ALU64 && opcode == BPF_ADD &&
-                   (dst_reg->type == PTR_TO_MAP_VALUE ||
-                    dst_reg->type == PTR_TO_MAP_VALUE_ADJ))
-                       dst_reg->type = PTR_TO_MAP_VALUE_ADJ;
-               else
-                       mark_reg_unknown_value(regs, insn->dst_reg);
+               return adjust_reg_min_max_vals(env, insn);
        }
 
        return 0;
@@ -2189,6 +2222,17 @@ static void find_good_pkt_pointers(struct bpf_verifier_state *state,
        struct bpf_reg_state *regs = state->regs, *reg;
        int i;
 
+       if (dst_reg->off < 0)
+               /* This doesn't give us any range */
+               return;
+
+       if (dst_reg->max_value > MAX_PACKET_OFF ||
+           dst_reg->max_value + dst_reg->off > MAX_PACKET_OFF)
+               /* Risk of overflow.  For instance, ptr + (1<<63) may be less
+                * than pkt_end, but that's because it's also less than pkt.
+                */
+               return;
+
        /* LLVM can generate two kind of checks:
         *
         * Type 1:
@@ -2219,30 +2263,44 @@ static void find_good_pkt_pointers(struct bpf_verifier_state *state,
         * so that range of bytes [r3, r3 + 8) is safe to access.
         */
 
+       /* If our ids match, then we must have the same max_value.  And we
+        * don't care about the other reg's fixed offset, since if it's too big
+        * the range won't allow anything.
+        * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
+        */
        for (i = 0; i < MAX_BPF_REG; i++)
                if (regs[i].type == PTR_TO_PACKET && regs[i].id == dst_reg->id)
                        /* keep the maximum range already checked */
-                       regs[i].range = max(regs[i].range, dst_reg->off);
+                       regs[i].range = max_t(u16, regs[i].range, dst_reg->off);
 
        for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
                if (state->stack_slot_type[i] != STACK_SPILL)
                        continue;
                reg = &state->spilled_regs[i / BPF_REG_SIZE];
                if (reg->type == PTR_TO_PACKET && reg->id == dst_reg->id)
-                       reg->range = max(reg->range, dst_reg->off);
+                       reg->range = max_t(u16, reg->range, dst_reg->off);
        }
 }
 
 /* Adjusts the register min/max values in the case that the dst_reg is the
  * variable register that we are working on, and src_reg is a constant or we're
  * simply doing a BPF_K check.
+ * In JEQ/JNE cases we also adjust the var_off values.
  */
 static void reg_set_min_max(struct bpf_reg_state *true_reg,
                            struct bpf_reg_state *false_reg, u64 val,
                            u8 opcode)
 {
        bool value_from_signed = true;
-       bool is_range = true;
+
+       /* If the dst_reg is a pointer, we can't learn anything about its
+        * variable offset from the compare (unless src_reg were a pointer into
+        * the same object, but we don't bother with that.
+        * Since false_reg and true_reg have the same type by construction, we
+        * only need to check one of them for pointerness.
+        */
+       if (__is_pointer_value(false, false_reg))
+               return;
 
        switch (opcode) {
        case BPF_JEQ:
@@ -2250,14 +2308,14 @@ static void reg_set_min_max(struct bpf_reg_state *true_reg,
                 * true then we know for sure.
                 */
                true_reg->max_value = true_reg->min_value = val;
-               is_range = false;
+               true_reg->var_off = tnum_const(val);
                break;
        case BPF_JNE:
                /* If this is true we know nothing Jon Snow, but if it is false
                 * we know the value for sure;
                 */
                false_reg->max_value = false_reg->min_value = val;
-               is_range = false;
+               false_reg->var_off = tnum_const(val);
                break;
        case BPF_JGT:
                value_from_signed = false;
@@ -2305,23 +2363,19 @@ static void reg_set_min_max(struct bpf_reg_state *true_reg,
 
        check_reg_overflow(false_reg);
        check_reg_overflow(true_reg);
-       if (is_range) {
-               if (__is_pointer_value(false, false_reg))
-                       reset_reg_range_values(false_reg, 0);
-               if (__is_pointer_value(false, true_reg))
-                       reset_reg_range_values(true_reg, 0);
-       }
 }
 
-/* Same as above, but for the case that dst_reg is a CONST_IMM reg and src_reg
- * is the variable reg.
+/* Same as above, but for the case that dst_reg holds a constant and src_reg is
+ * the variable reg.
  */
 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
                                struct bpf_reg_state *false_reg, u64 val,
                                u8 opcode)
 {
        bool value_from_signed = true;
-       bool is_range = true;
+
+       if (__is_pointer_value(false, false_reg))
+               return;
 
        switch (opcode) {
        case BPF_JEQ:
@@ -2329,14 +2383,14 @@ static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
                 * true then we know for sure.
                 */
                true_reg->max_value = true_reg->min_value = val;
-               is_range = false;
+               true_reg->var_off = tnum_const(val);
                break;
        case BPF_JNE:
                /* If this is true we know nothing Jon Snow, but if it is false
                 * we know the value for sure;
                 */
                false_reg->max_value = false_reg->min_value = val;
-               is_range = false;
+               false_reg->var_off = tnum_const(val);
                break;
        case BPF_JGT:
                value_from_signed = false;
@@ -2385,27 +2439,60 @@ static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
 
        check_reg_overflow(false_reg);
        check_reg_overflow(true_reg);
-       if (is_range) {
-               if (__is_pointer_value(false, false_reg))
-                       reset_reg_range_values(false_reg, 0);
-               if (__is_pointer_value(false, true_reg))
-                       reset_reg_range_values(true_reg, 0);
+}
+
+/* Regs are known to be equal, so intersect their min/max/var_off */
+static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
+                                 struct bpf_reg_state *dst_reg)
+{
+       src_reg->min_value = dst_reg->min_value = max(src_reg->min_value,
+                                                     dst_reg->min_value);
+       src_reg->max_value = dst_reg->max_value = min(src_reg->max_value,
+                                                     dst_reg->max_value);
+       src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
+                                                            dst_reg->var_off);
+       check_reg_overflow(src_reg);
+       check_reg_overflow(dst_reg);
+}
+
+static void reg_combine_min_max(struct bpf_reg_state *true_src,
+                               struct bpf_reg_state *true_dst,
+                               struct bpf_reg_state *false_src,
+                               struct bpf_reg_state *false_dst,
+                               u8 opcode)
+{
+       switch (opcode) {
+       case BPF_JEQ:
+               __reg_combine_min_max(true_src, true_dst);
+               break;
+       case BPF_JNE:
+               __reg_combine_min_max(false_src, false_dst);
        }
 }
 
 static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
-                        enum bpf_reg_type type)
+                        bool is_null)
 {
        struct bpf_reg_state *reg = &regs[regno];
 
        if (reg->type == PTR_TO_MAP_VALUE_OR_NULL && reg->id == id) {
-               if (type == UNKNOWN_VALUE) {
-                       __mark_reg_unknown_value(regs, regno);
+               /* Old offset (both fixed and variable parts) should
+                * have been known-zero, because we don't allow pointer
+                * arithmetic on pointers that might be NULL.
+                */
+               if (WARN_ON_ONCE(reg->min_value || reg->max_value ||
+                                reg->var_off.value || reg->var_off.mask ||
+                                reg->off)) {
+                       reg->min_value = reg->max_value = reg->off = 0;
+                       reg->var_off = tnum_const(0);
+               }
+               if (is_null) {
+                       reg->type = SCALAR_VALUE;
                } else if (reg->map_ptr->inner_map_meta) {
                        reg->type = CONST_PTR_TO_MAP;
                        reg->map_ptr = reg->map_ptr->inner_map_meta;
                } else {
-                       reg->type = type;
+                       reg->type = PTR_TO_MAP_VALUE;
                }
                /* We don't need id from this point onwards anymore, thus we
                 * should better reset it, so that state pruning has chances
@@ -2419,19 +2506,19 @@ static void mark_map_reg(struct bpf_reg_state *regs, u32 regno, u32 id,
  * be folded together at some point.
  */
 static void mark_map_regs(struct bpf_verifier_state *state, u32 regno,
-                         enum bpf_reg_type type)
+                         bool is_null)
 {
        struct bpf_reg_state *regs = state->regs;
        u32 id = regs[regno].id;
        int i;
 
        for (i = 0; i < MAX_BPF_REG; i++)
-               mark_map_reg(regs, i, id, type);
+               mark_map_reg(regs, i, id, is_null);
 
        for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
                if (state->stack_slot_type[i] != STACK_SPILL)
                        continue;
-               mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, type);
+               mark_map_reg(state->spilled_regs, i / BPF_REG_SIZE, id, is_null);
        }
 }
 
@@ -2481,7 +2568,8 @@ static int check_cond_jmp_op(struct bpf_verifier_env *env,
        /* detect if R == 0 where R was initialized to zero earlier */
        if (BPF_SRC(insn->code) == BPF_K &&
            (opcode == BPF_JEQ || opcode == BPF_JNE) &&
-           dst_reg->type == CONST_IMM && dst_reg->imm == insn->imm) {
+           dst_reg->type == SCALAR_VALUE &&
+           tnum_equals_const(dst_reg->var_off, insn->imm)) {
                if (opcode == BPF_JEQ) {
                        /* if (imm == imm) goto pc+off;
                         * only follow the goto, ignore fall-through
@@ -2503,17 +2591,30 @@ static int check_cond_jmp_op(struct bpf_verifier_env *env,
 
        /* detect if we are comparing against a constant value so we can adjust
         * our min/max values for our dst register.
+        * this is only legit if both are scalars (or pointers to the same
+        * object, I suppose, but we don't support that right now), because
+        * otherwise the different base pointers mean the offsets aren't
+        * comparable.
         */
        if (BPF_SRC(insn->code) == BPF_X) {
-               if (regs[insn->src_reg].type == CONST_IMM)
-                       reg_set_min_max(&other_branch->regs[insn->dst_reg],
-                                       dst_reg, regs[insn->src_reg].imm,
-                                       opcode);
-               else if (dst_reg->type == CONST_IMM)
-                       reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
-                                           &regs[insn->src_reg], dst_reg->imm,
-                                           opcode);
-       } else {
+               if (dst_reg->type == SCALAR_VALUE &&
+                   regs[insn->src_reg].type == SCALAR_VALUE) {
+                       if (tnum_is_const(regs[insn->src_reg].var_off))
+                               reg_set_min_max(&other_branch->regs[insn->dst_reg],
+                                               dst_reg, regs[insn->src_reg].var_off.value,
+                                               opcode);
+                       else if (tnum_is_const(dst_reg->var_off))
+                               reg_set_min_max_inv(&other_branch->regs[insn->src_reg],
+                                                   &regs[insn->src_reg],
+                                                   dst_reg->var_off.value, opcode);
+                       else if (opcode == BPF_JEQ || opcode == BPF_JNE)
+                               /* Comparing for equality, we can combine knowledge */
+                               reg_combine_min_max(&other_branch->regs[insn->src_reg],
+                                                   &other_branch->regs[insn->dst_reg],
+                                                   &regs[insn->src_reg],
+                                                   &regs[insn->dst_reg], opcode);
+               }
+       } else if (dst_reg->type == SCALAR_VALUE) {
                reg_set_min_max(&other_branch->regs[insn->dst_reg],
                                        dst_reg, insn->imm, opcode);
        }
@@ -2525,10 +2626,8 @@ static int check_cond_jmp_op(struct bpf_verifier_env *env,
                /* Mark all identical map registers in each branch as either
                 * safe or unknown depending R == 0 or R != 0 conditional.
                 */
-               mark_map_regs(this_branch, insn->dst_reg,
-                             opcode == BPF_JEQ ? PTR_TO_MAP_VALUE : UNKNOWN_VALUE);
-               mark_map_regs(other_branch, insn->dst_reg,
-                             opcode == BPF_JEQ ? UNKNOWN_VALUE : PTR_TO_MAP_VALUE);
+               mark_map_regs(this_branch, insn->dst_reg, opcode == BPF_JNE);
+               mark_map_regs(other_branch, insn->dst_reg, opcode == BPF_JEQ);
        } else if (BPF_SRC(insn->code) == BPF_X && opcode == BPF_JGT &&
                   dst_reg->type == PTR_TO_PACKET &&
                   regs[insn->src_reg].type == PTR_TO_PACKET_END) {
@@ -2576,8 +2675,11 @@ static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
        if (insn->src_reg == 0) {
                u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
 
-               regs[insn->dst_reg].type = CONST_IMM;
-               regs[insn->dst_reg].imm = imm;
+               regs[insn->dst_reg].type = SCALAR_VALUE;
+               regs[insn->dst_reg].min_value = imm;
+               regs[insn->dst_reg].max_value = imm;
+               check_reg_overflow(&regs[insn->dst_reg]);
+               regs[insn->dst_reg].var_off = tnum_const(imm);
                regs[insn->dst_reg].id = 0;
                return 0;
        }
@@ -2659,7 +2761,7 @@ static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
        /* mark destination R0 register as readable, since it contains
         * the value fetched from the packet
         */
-       regs[BPF_REG_0].type = UNKNOWN_VALUE;
+       mark_reg_unknown(regs, BPF_REG_0);
        return 0;
 }
 
@@ -2862,57 +2964,145 @@ err_free:
        return ret;
 }
 
-/* the following conditions reduce the number of explored insns
- * from ~140k to ~80k for ultra large programs that use a lot of ptr_to_packet
+/* check %cur's range satisfies %old's */
+static bool range_within(struct bpf_reg_state *old,
+                        struct bpf_reg_state *cur)
+{
+       return old->min_value <= cur->min_value &&
+              old->max_value >= cur->max_value;
+}
+
+/* Maximum number of register states that can exist at once */
+#define ID_MAP_SIZE    (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
+struct idpair {
+       u32 old;
+       u32 cur;
+};
+
+/* If in the old state two registers had the same id, then they need to have
+ * the same id in the new state as well.  But that id could be different from
+ * the old state, so we need to track the mapping from old to new ids.
+ * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
+ * regs with old id 5 must also have new id 9 for the new state to be safe.  But
+ * regs with a different old id could still have new id 9, we don't care about
+ * that.
+ * So we look through our idmap to see if this old id has been seen before.  If
+ * so, we require the new id to match; otherwise, we add the id pair to the map.
  */
-static bool compare_ptrs_to_packet(struct bpf_verifier_env *env,
-                                  struct bpf_reg_state *old,
-                                  struct bpf_reg_state *cur)
+static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
 {
-       if (old->id != cur->id)
-               return false;
+       unsigned int i;
+
+       for (i = 0; i < ID_MAP_SIZE; i++) {
+               if (!idmap[i].old) {
+                       /* Reached an empty slot; haven't seen this id before */
+                       idmap[i].old = old_id;
+                       idmap[i].cur = cur_id;
+                       return true;
+               }
+               if (idmap[i].old == old_id)
+                       return idmap[i].cur == cur_id;
+       }
+       /* We ran out of idmap slots, which should be impossible */
+       WARN_ON_ONCE(1);
+       return false;
+}
 
-       /* old ptr_to_packet is more conservative, since it allows smaller
-        * range. Ex:
-        * old(off=0,r=10) is equal to cur(off=0,r=20), because
-        * old(off=0,r=10) means that with range=10 the verifier proceeded
-        * further and found no issues with the program. Now we're in the same
-        * spot with cur(off=0,r=20), so we're safe too, since anything further
-        * will only be looking at most 10 bytes after this pointer.
-        */
-       if (old->off == cur->off && old->range < cur->range)
+/* Returns true if (rold safe implies rcur safe) */
+static bool regsafe(struct bpf_reg_state *rold,
+                   struct bpf_reg_state *rcur,
+                   bool varlen_map_access, struct idpair *idmap)
+{
+       if (memcmp(rold, rcur, sizeof(*rold)) == 0)
                return true;
 
-       /* old(off=20,r=10) is equal to cur(off=22,re=22 or 5 or 0)
-        * since both cannot be used for packet access and safe(old)
-        * pointer has smaller off that could be used for further
-        * 'if (ptr > data_end)' check
-        * Ex:
-        * old(off=20,r=10) and cur(off=22,r=22) and cur(off=22,r=0) mean
-        * that we cannot access the packet.
-        * The safe range is:
-        * [ptr, ptr + range - off)
-        * so whenever off >=range, it means no safe bytes from this pointer.
-        * When comparing old->off <= cur->off, it means that older code
-        * went with smaller offset and that offset was later
-        * used to figure out the safe range after 'if (ptr > data_end)' check
-        * Say, 'old' state was explored like:
-        * ... R3(off=0, r=0)
-        * R4 = R3 + 20
-        * ... now R4(off=20,r=0)  <-- here
-        * if (R4 > data_end)
-        * ... R4(off=20,r=20), R3(off=0,r=20) and R3 can be used to access.
-        * ... the code further went all the way to bpf_exit.
-        * Now the 'cur' state at the mark 'here' has R4(off=30,r=0).
-        * old_R4(off=20,r=0) equal to cur_R4(off=30,r=0), since if the verifier
-        * goes further, such cur_R4 will give larger safe packet range after
-        * 'if (R4 > data_end)' and all further insn were already good with r=20,
-        * so they will be good with r=30 and we can prune the search.
-        */
-       if (!env->strict_alignment && old->off <= cur->off &&
-           old->off >= old->range && cur->off >= cur->range)
+       if (rold->type == NOT_INIT)
+               /* explored state can't have used this */
                return true;
+       if (rcur->type == NOT_INIT)
+               return false;
+       switch (rold->type) {
+       case SCALAR_VALUE:
+               if (rcur->type == SCALAR_VALUE) {
+                       /* new val must satisfy old val knowledge */
+                       return range_within(rold, rcur) &&
+                              tnum_in(rold->var_off, rcur->var_off);
+               } else {
+                       /* if we knew anything about the old value, we're not
+                        * equal, because we can't know anything about the
+                        * scalar value of the pointer in the new value.
+                        */
+                       return rold->min_value == BPF_REGISTER_MIN_RANGE &&
+                              rold->max_value == BPF_REGISTER_MAX_RANGE &&
+                              tnum_is_unknown(rold->var_off);
+               }
+       case PTR_TO_MAP_VALUE:
+               if (varlen_map_access) {
+                       /* If the new min/max/var_off satisfy the old ones and
+                        * everything else matches, we are OK.
+                        * We don't care about the 'id' value, because nothing
+                        * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
+                        */
+                       return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
+                              range_within(rold, rcur) &&
+                              tnum_in(rold->var_off, rcur->var_off);
+               } else {
+                       /* If the ranges/var_off were not the same, but
+                        * everything else was and we didn't do a variable
+                        * access into a map then we are a-ok.
+                        */
+                       return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0;
+               }
+       case PTR_TO_MAP_VALUE_OR_NULL:
+               /* a PTR_TO_MAP_VALUE could be safe to use as a
+                * PTR_TO_MAP_VALUE_OR_NULL into the same map.
+                * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
+                * checked, doing so could have affected others with the same
+                * id, and we can't check for that because we lost the id when
+                * we converted to a PTR_TO_MAP_VALUE.
+                */
+               if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
+                       return false;
+               if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
+                       return false;
+               /* Check our ids match any regs they're supposed to */
+               return check_ids(rold->id, rcur->id, idmap);
+       case PTR_TO_PACKET:
+               if (rcur->type != PTR_TO_PACKET)
+                       return false;
+               /* We must have at least as much range as the old ptr
+                * did, so that any accesses which were safe before are
+                * still safe.  This is true even if old range < old off,
+                * since someone could have accessed through (ptr - k), or
+                * even done ptr -= k in a register, to get a safe access.
+                */
+               if (rold->range > rcur->range)
+                       return false;
+               /* If the offsets don't match, we can't trust our alignment;
+                * nor can we be sure that we won't fall out of range.
+                */
+               if (rold->off != rcur->off)
+                       return false;
+               /* id relations must be preserved */
+               if (rold->id && !check_ids(rold->id, rcur->id, idmap))
+                       return false;
+               /* new val must satisfy old val knowledge */
+               return range_within(rold, rcur) &&
+                      tnum_in(rold->var_off, rcur->var_off);
+       case PTR_TO_CTX:
+       case CONST_PTR_TO_MAP:
+       case PTR_TO_STACK:
+       case PTR_TO_PACKET_END:
+               /* Only valid matches are exact, which memcmp() above
+                * would have accepted
+                */
+       default:
+               /* Don't know what's going on, just say it's not safe */
+               return false;
+       }
 
+       /* Shouldn't get here; if we do, say it's not safe */
+       WARN_ON_ONCE(1);
        return false;
 }
 
@@ -2947,43 +3137,19 @@ static bool states_equal(struct bpf_verifier_env *env,
                         struct bpf_verifier_state *cur)
 {
        bool varlen_map_access = env->varlen_map_value_access;
-       struct bpf_reg_state *rold, *rcur;
+       struct idpair *idmap;
+       bool ret = false;
        int i;
 
-       for (i = 0; i < MAX_BPF_REG; i++) {
-               rold = &old->regs[i];
-               rcur = &cur->regs[i];
-
-               if (memcmp(rold, rcur, sizeof(*rold)) == 0)
-                       continue;
-
-               /* If the ranges were not the same, but everything else was and
-                * we didn't do a variable access into a map then we are a-ok.
-                */
-               if (!varlen_map_access &&
-                   memcmp(rold, rcur, offsetofend(struct bpf_reg_state, id)) == 0)
-                       continue;
-
-               /* If we didn't map access then again we don't care about the
-                * mismatched range values and it's ok if our old type was
-                * UNKNOWN and we didn't go to a NOT_INIT'ed reg.
-                */
-               if (rold->type == NOT_INIT ||
-                   (!varlen_map_access && rold->type == UNKNOWN_VALUE &&
-                    rcur->type != NOT_INIT))
-                       continue;
-
-               /* Don't care about the reg->id in this case. */
-               if (rold->type == PTR_TO_MAP_VALUE_OR_NULL &&
-                   rcur->type == PTR_TO_MAP_VALUE_OR_NULL &&
-                   rold->map_ptr == rcur->map_ptr)
-                       continue;
-
-               if (rold->type == PTR_TO_PACKET && rcur->type == PTR_TO_PACKET &&
-                   compare_ptrs_to_packet(env, rold, rcur))
-                       continue;
-
+       idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
+       /* If we failed to allocate the idmap, just say it's not safe */
+       if (!idmap)
                return false;
+
+       for (i = 0; i < MAX_BPF_REG; i++) {
+               if (!regsafe(&old->regs[i], &cur->regs[i], varlen_map_access,
+                            idmap))
+                       goto out_free;
        }
 
        for (i = 0; i < MAX_BPF_STACK; i++) {
@@ -2995,29 +3161,32 @@ static bool states_equal(struct bpf_verifier_env *env,
                         * this verifier states are not equivalent,
                         * return false to continue verification of this path
                         */
-                       return false;
+                       goto out_free;
                if (i % BPF_REG_SIZE)
                        continue;
                if (old->stack_slot_type[i] != STACK_SPILL)
                        continue;
-               if (memcmp(&old->spilled_regs[i / BPF_REG_SIZE],
-                          &cur->spilled_regs[i / BPF_REG_SIZE],
-                          sizeof(old->spilled_regs[0])))
-                       /* when explored and current stack slot types are
-                        * the same, check that stored pointers types
+               if (!regsafe(&old->spilled_regs[i / BPF_REG_SIZE],
+                            &cur->spilled_regs[i / BPF_REG_SIZE],
+                            varlen_map_access, idmap))
+                       /* when explored and current stack slot are both storing
+                        * spilled registers, check that stored pointers types
                         * are the same as well.
                         * Ex: explored safe path could have stored
-                        * (bpf_reg_state) {.type = PTR_TO_STACK, .imm = -8}
+                        * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
                         * but current path has stored:
-                        * (bpf_reg_state) {.type = PTR_TO_STACK, .imm = -16}
+                        * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
                         * such verifier states are not equivalent.
                         * return false to continue verification of this path
                         */
-                       return false;
+                       goto out_free;
                else
                        continue;
        }
-       return true;
+       ret = true;
+out_free:
+       kfree(idmap);
+       return ret;
 }
 
 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
@@ -3331,7 +3500,6 @@ process_bpf_exit:
                                verbose("invalid BPF_LD mode\n");
                                return -EINVAL;
                        }
-                       reset_reg_range_values(regs, insn->dst_reg);
                } else {
                        verbose("unknown insn class %d\n", class);
                        return -EINVAL;