lib/mpi: BUG: sleeping function called from invalid context on next-20160726

Hi,

with linux-next-20160726, I get this:

 BUG: sleeping function called from invalid context at /mnt/scratch/nic/linux-next/mm/slab.h:388
 in_atomic(): 1, irqs_disabled(): 0, pid: 369, name: systemd-udevd
 no locks held by systemd-udevd/369.
 CPU: 2 PID: 369 Comm: systemd-udevd Not tainted 4.7.0-rc1+ #248
 Hardware name: Dell Inc. Latitude E6540/0725FP, BIOS A10 06/26/2014
  0000000000000286 00000000899a9b52 ffff88003f53b8f8 ffffffff814472d5
  ffff8800c0752680 ffffffff81c557d8 ffff88003f53b920 ffffffff810dfba9
  ffffffff81c557d8 0000000000000184 0000000000000000 ffff88003f53b948
 Call Trace:
  [<ffffffff814472d5>] dump_stack+0x86/0xc1
  [<ffffffff810dfba9>] ___might_sleep+0x179/0x230
  [<ffffffff810dfca9>] __might_sleep+0x49/0x80
  [<ffffffff8125f1f1>] kmem_cache_alloc_trace+0x1d1/0x2e0
  [<ffffffff81479b20>] ? mpi_alloc+0x20/0x80
  [<ffffffff81479b20>] mpi_alloc+0x20/0x80
  [<ffffffff81477475>] mpi_read_raw_from_sgl+0xd5/0x1e0
  [<ffffffff813e99f6>] rsa_verify+0x66/0x100
  [<ffffffff813ea1be>] pkcs1pad_verify+0xae/0xf0
  [<ffffffff81404889>] public_key_verify_signature+0x1f9/0x290
  [<ffffffff81404935>] public_key_verify_signature_2+0x15/0x20
  [<ffffffff8140458c>] verify_signature+0x3c/0x50
  [<ffffffff8140680d>] pkcs7_validate_trust+0x11d/0x230
  [<ffffffff811eb132>] verify_pkcs7_signature+0xa2/0x150
  [<ffffffff8115747d>] mod_verify_sig+0xdd/0x130
  [<ffffffff811541cc>] load_module+0x16c/0x2970
  [<ffffffff8128b95b>] ? vfs_read+0x11b/0x130
  [<ffffffff81292342>] ? kernel_read_file+0x152/0x170
  [<ffffffff81156c66>] SYSC_finit_module+0xe6/0x120
  [<ffffffff81156cbe>] SyS_finit_module+0xe/0x10
  [<ffffffff81003fe7>] do_syscall_64+0x67/0x190
  [<ffffffff8189ab3f>] entry_SYSCALL64_slow_path+0x25/0x25

Reason is 127827b9c295 ("lib/mpi: Do not do sg_virt") which makes mpi_read_raw_from_sgl() calling mpi_alloc() while having a sg entry mapped via kmap_atomic() and thus, preemption disabled.

I would have sent a patch, but there is another point which puzzles me in mpi_read_raw_from_sgl():

  [...]
  const u8 *buff;
  [...]
  sg_miter_start(&miter, sgl, ents, SG_MITER_ATOMIC | SG_MITER_FROM_SG);

  lzeros = 0;
  len = 0;
  while (nbytes > 0) {
  	while (len && !*buff) {
  		lzeros++;
  		len--;
  		buff++;
  	}

Thus, buff isn't initialized before its first use? Or am I misreading something here?

Thanks,

Nicolai

Does this patch help?

On the first entry len is zero therefore we will go to the end of the loop and initialise buff. Anyway, it will no longer be as confusing with this patch applied.

Thanks,

---8<---
Subject: lib/mpi: Fix SG miter leak

In mpi_read_raw_from_sgl we may leak the SG miter resouces after reading the leading zeroes. This patch fixes this by stopping the iteration once the leading zeroes have been read.

index c6272ae..5a0f75a 100644
@@ -363,6 +363,9 @@ MPI mpi_read_raw_from_sgl(struct scatterlist *sgl, unsigned int nbytes)
 		lzeros = 0;
 	}
 
+	miter.consumed = lzeros;
+	sg_miter_stop(&miter);
+
 	nbytes -= lzeros;
 	nbits = nbytes * 8;
 	if (nbits > MAX_EXTERN_MPI_BITS) {

@@ -390,7 +393,10 @@ MPI mpi_read_raw_from_sgl(struct scatterlist *sgl, unsigned int nbytes)
 	z = BYTES_PER_MPI_LIMB - nbytes % BYTES_PER_MPI_LIMB;
 	z %= BYTES_PER_MPI_LIMB;
 
-	for (;;) {
+	while (sg_miter_next(&miter)) {
+		buff = miter.addr;
+		len = miter.length;
+
 		for (x = 0; x < len; x++) {
 			a <<= 8;
 			a |= *buff++;

@@ -400,12 +406,6 @@ MPI mpi_read_raw_from_sgl(struct scatterlist *sgl, unsigned int nbytes)
 			}
 		}
 		z += x;
-
-		if (!sg_miter_next(&miter))
-			break;
-
-		buff = miter.addr;
-		len = miter.length;
 	}
 
 	return val;

Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Yes, works like a charm now!

Hah! Thanks, although being obvious, I didn't see this...

Thanks,

Nicolai