drbd: fix potential distributed deadlock

We limit ourselves to a configurable maximum number of pages used as
temporary bio pages.

If the configured "max_buffers" is not big enough to match the bandwidth
of the respective deployment, a distributed deadlock could be triggered
by e.g. fast online verify and heavy application IO.

TCP connections would block on congestion, because both receivers
would wait on pages to become available.

Fortunately the respective senders in this case would be able to give
back some pages already. So do that.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

1 files changed, 59 insertions, 35 deletions

diff --git a/drivers/block/drbd/drbd_worker.c b/drivers/block/drbd/drbd_worker.c
index f7e6c92f8d0..b5e53695fd7 100644
--- a/drivers/block/drbd/drbd_worker.c
+++ b/drivers/block/drbd/drbd_worker.c
@@ -297,42 +297,48 @@ void drbd_csum_bio(struct drbd_conf *mdev, struct crypto_hash *tfm, struct bio *
         crypto_hash_final(&desc, digest);
 }
 
-static int w_e_send_csum(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
+/* TODO merge common code with w_e_end_ov_req */
+int w_e_send_csum(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
 {
         struct drbd_epoch_entry *e = container_of(w, struct drbd_epoch_entry, w);
         int digest_size;
         void *digest;
-        int ok;
+        int ok = 1;
 
         D_ASSERT(e->block_id == DRBD_MAGIC + 0xbeef);
 
-        if (unlikely(cancel)) {
-                drbd_free_ee(mdev, e);
-                return 1;
-        }
+        if (unlikely(cancel))
+                goto out;
 
-        if (likely((e->flags & EE_WAS_ERROR) == 0)) {
-                digest_size = crypto_hash_digestsize(mdev->csums_tfm);
-                digest = kmalloc(digest_size, GFP_NOIO);
-                if (digest) {
-                        drbd_csum_ee(mdev, mdev->csums_tfm, e, digest);
+        if (likely((e->flags & EE_WAS_ERROR) != 0))
+                goto out;
 
-                        inc_rs_pending(mdev);
-                        ok = drbd_send_drequest_csum(mdev,
-                                                     e->sector,
-                                                     e->size,
-                                                     digest,
-                                                     digest_size,
-                                                     P_CSUM_RS_REQUEST);
-                        kfree(digest);
-                } else {
-                        dev_err(DEV, "kmalloc() of digest failed.\n");
-                        ok = 0;
-                }
-        } else
-                ok = 1;
+        digest_size = crypto_hash_digestsize(mdev->csums_tfm);
+        digest = kmalloc(digest_size, GFP_NOIO);
+        if (digest) {
+                sector_t sector = e->sector;
+                unsigned int size = e->size;
+                drbd_csum_ee(mdev, mdev->csums_tfm, e, digest);
+                /* Free e and pages before send.
+                 * In case we block on congestion, we could otherwise run into
+                 * some distributed deadlock, if the other side blocks on
+                 * congestion as well, because our receiver blocks in
+                 * drbd_pp_alloc due to pp_in_use > max_buffers. */
+                drbd_free_ee(mdev, e);
+                e = NULL;
+                inc_rs_pending(mdev);
+                ok = drbd_send_drequest_csum(mdev, sector, size,
+                                             digest, digest_size,
+                                             P_CSUM_RS_REQUEST);
+                kfree(digest);
+        } else {
+                dev_err(DEV, "kmalloc() of digest failed.\n");
+                ok = 0;
+        }
 
-        drbd_free_ee(mdev, e);
+out:
+        if (e)
+                drbd_free_ee(mdev, e);
 
         if (unlikely(!ok))
                 dev_err(DEV, "drbd_send_drequest(..., csum) failed\n");
@@ -1071,9 +1077,12 @@ int w_e_end_csum_rs_req(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
         return ok;
 }
 
+/* TODO merge common code with w_e_send_csum */
 int w_e_end_ov_req(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
 {
         struct drbd_epoch_entry *e = container_of(w, struct drbd_epoch_entry, w);
+        sector_t sector = e->sector;
+        unsigned int size = e->size;
         int digest_size;
         void *digest;
         int ok = 1;
@@ -1093,17 +1102,25 @@ int w_e_end_ov_req(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
         else
                 memset(digest, 0, digest_size);
 
+        /* Free e and pages before send.
+         * In case we block on congestion, we could otherwise run into
+         * some distributed deadlock, if the other side blocks on
+         * congestion as well, because our receiver blocks in
+         * drbd_pp_alloc due to pp_in_use > max_buffers. */
+        drbd_free_ee(mdev, e);
+        e = NULL;
         inc_rs_pending(mdev);
-        ok = drbd_send_drequest_csum(mdev, e->sector, e->size,
-                                     digest, digest_size, P_OV_REPLY);
+        ok = drbd_send_drequest_csum(mdev, sector, size,
+                                     digest, digest_size,
+                                     P_OV_REPLY);
         if (!ok)
                 dec_rs_pending(mdev);
         kfree(digest);
 
 out:
-        drbd_free_ee(mdev, e);
+        if (e)
+                drbd_free_ee(mdev, e);
         dec_unacked(mdev);
-
         return ok;
 }
 
@@ -1122,8 +1139,10 @@ int w_e_end_ov_reply(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
 {
         struct drbd_epoch_entry *e = container_of(w, struct drbd_epoch_entry, w);
         struct digest_info *di;
-        int digest_size;
         void *digest;
+        sector_t sector = e->sector;
+        unsigned int size = e->size;
+        int digest_size;
         int ok, eq = 0;
 
         if (unlikely(cancel)) {
@@ -1153,16 +1172,21 @@ int w_e_end_ov_reply(struct drbd_conf *mdev, struct drbd_work *w, int cancel)
                 }
         }
 
-        dec_unacked(mdev);
+                /* Free e and pages before send.
+                 * In case we block on congestion, we could otherwise run into
+                 * some distributed deadlock, if the other side blocks on
+                 * congestion as well, because our receiver blocks in
+                 * drbd_pp_alloc due to pp_in_use > max_buffers. */
+        drbd_free_ee(mdev, e);
         if (!eq)
-                drbd_ov_oos_found(mdev, e->sector, e->size);
+                drbd_ov_oos_found(mdev, sector, size);
         else
                 ov_oos_print(mdev);
 
-        ok = drbd_send_ack_ex(mdev, P_OV_RESULT, e->sector, e->size,
+        ok = drbd_send_ack_ex(mdev, P_OV_RESULT, sector, size,
                               eq ? ID_IN_SYNC : ID_OUT_OF_SYNC);
 
-        drbd_free_ee(mdev, e);
+        dec_unacked(mdev);
 
         --mdev->ov_left;