From: Theodore Ts'o Date: Sun, 15 Jun 2014 20:59:24 +0000 (-0400) Subject: random: use registers from interrupted code for CPU's w/o a cycle counter X-Git-Url: https://git.karo-electronics.de/?a=commitdiff_plain;h=ee3e00e9e7101c80a2ff2d5672d4b486bf001b88;p=linux-beck.git random: use registers from interrupted code for CPU's w/o a cycle counter For CPU's that don't have a cycle counter, or something equivalent which can be used for random_get_entropy(), random_get_entropy() will always return 0. In that case, substitute with the saved interrupt registers to add a bit more unpredictability. Some folks have suggested hashing all of the registers unconditionally, but this would increase the overhead of add_interrupt_randomness() by at least an order of magnitude, and this would very likely be unacceptable. The changes in this commit have been benchmarked as mostly unaffecting the overhead of add_interrupt_randomness() if the entropy counter is present, and doubling the overhead if it is not present. Signed-off-by: Theodore Ts'o Cc: Jörn Engel --- diff --git a/drivers/char/random.c b/drivers/char/random.c index 914b1575df8f..aa22fe551c2a 100644 --- a/drivers/char/random.c +++ b/drivers/char/random.c @@ -551,9 +551,8 @@ static void mix_pool_bytes(struct entropy_store *r, const void *in, struct fast_pool { __u32 pool[4]; unsigned long last; + unsigned short reg_idx; unsigned char count; - unsigned char notimer_count; - unsigned char rotate; }; /* @@ -857,6 +856,17 @@ static void add_interrupt_bench(cycles_t start) #define add_interrupt_bench(x) #endif +static __u32 get_reg(struct fast_pool *f, struct pt_regs *regs) +{ + __u32 *ptr = (__u32 *) regs; + + if (regs == NULL) + return 0; + if (f->reg_idx >= sizeof(struct pt_regs) / sizeof(__u32)) + f->reg_idx = 0; + return *(ptr + f->reg_idx++); +} + void add_interrupt_randomness(int irq, int irq_flags) { struct entropy_store *r; @@ -869,28 +879,23 @@ void add_interrupt_randomness(int irq, int irq_flags) unsigned long seed; int credit = 0; + if (cycles == 0) + cycles = get_reg(fast_pool, regs); c_high = (sizeof(cycles) > 4) ? cycles >> 32 : 0; j_high = (sizeof(now) > 4) ? now >> 32 : 0; fast_pool->pool[0] ^= cycles ^ j_high ^ irq; fast_pool->pool[1] ^= now ^ c_high; ip = regs ? instruction_pointer(regs) : _RET_IP_; fast_pool->pool[2] ^= ip; - fast_pool->pool[3] ^= ip >> 32; + fast_pool->pool[3] ^= (sizeof(ip) > 4) ? ip >> 32 : + get_reg(fast_pool, regs); fast_mix(fast_pool); - if ((irq_flags & __IRQF_TIMER) == 0) - fast_pool->notimer_count++; add_interrupt_bench(cycles); - if (cycles) { - if ((fast_pool->count < 64) && - !time_after(now, fast_pool->last + HZ)) - return; - } else { - /* CPU does not have a cycle counting register :-( */ - if (fast_pool->count < 64) - return; - } + if ((fast_pool->count < 64) && + !time_after(now, fast_pool->last + HZ)) + return; r = nonblocking_pool.initialized ? &input_pool : &nonblocking_pool; if (!spin_trylock(&r->lock)) @@ -910,18 +915,10 @@ void add_interrupt_randomness(int irq, int irq_flags) } spin_unlock(&r->lock); - /* - * If we have a valid cycle counter or if the majority of - * interrupts collected were non-timer interrupts, then give - * an entropy credit of 1 bit. Yes, this is being very - * conservative. - */ - if (cycles || (fast_pool->notimer_count >= 32)) - credit++; - - fast_pool->count = fast_pool->notimer_count = 0; + fast_pool->count = 0; - credit_entropy_bits(r, credit); + /* award one bit for the contents of the fast pool */ + credit_entropy_bits(r, credit + 1); } #ifdef CONFIG_BLOCK