Skip to contentSkip to conversation

hw/net/tulip: Fix LP#1874539

RFC PATCH In review Philippe Mathieu-Daudé 5 participants 13 messages Apr 23, 2020 – May 13, 2020

lore
Share & tools
lore
#0 lore raw

cover hw/net/tulip: Fix LP#1874539

Attempt to fix the launchpad bug filled by Helge:

In a qemu-system-hppa system, qemu release v5.0.0-rc, the tulip nic driver is broken. The tulip nic is detected, even getting DHCP info does work. But when trying to download bigger files via network, the tulip driver gets stuck.

Philippe Mathieu-Daudé (3):
hw/net/tulip: Fix 'Descriptor Error' definition hw/net/tulip: Log descriptor overflows
hw/net/tulip: Set descriptor error bit when lenght is incorrect

 hw/net/tulip.h |  2 +-
 hw/net/tulip.c | 32 ++++++++++++++++++++++++++++----
 2 files changed, 29 insertions(+), 5 deletions(-)
Signature · 1 line
2.21.1
#1 lore raw

RFC PATCH 1/3 hw/net/tulip: Fix 'Descriptor Error' definition

Bit #14 is "DE" for 'Descriptor Error':

When set, indicates a frame truncation caused by a frame that does not fit within the current descriptor buffers, and that the 21143 does not own the next descriptor.

[Table 4-1. RDES0 Bit Fields Description]

---
 hw/net/tulip.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
hw/net/tulip.h +1 -1
index 97521b21db..5271aad8d5 100644
@@ -211,7 +211,7 @@
 #define RDES0_RF         BIT(11)
 #define RDES0_DT_SHIFT   12
 #define RDES0_DT_MASK    3
-#define RDES0_LE         BIT(14)
+#define RDES0_DE         BIT(14)
 #define RDES0_ES         BIT(15)
 #define RDES0_FL_SHIFT   16
 #define RDES0_FL_MASK    0x3fff
Signature · 1 line
2.21.1
#2 lore raw

RFC PATCH 2/3 hw/net/tulip: Log descriptor overflows

In reply to Philippe Mathieu-Daudé

Log with GUEST_ERROR what the guest is doing wrong.

---
 hw/net/tulip.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
hw/net/tulip.c +16 -0
index 1295f51d07..470f635acb 100644
@@ -172,6 +172,11 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
         }
 
         if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
+            qemu_log_mask(LOG_GUEST_ERROR,
+                          "%s: descriptor overflow "
+                          "(ofs: %u, len:%d, size:%zu)\n",
+                          __func__, s->rx_frame_len, len,
+                          sizeof(s->rx_frame));
             return;
         }
         pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame +
@@ -187,6 +192,11 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
         }
 
         if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
+            qemu_log_mask(LOG_GUEST_ERROR,
+                          "%s: descriptor overflow "
+                          "(ofs: %u, len:%d, size:%zu)\n",
+                          __func__, s->rx_frame_len, len,
+                          sizeof(s->rx_frame));
             return;
         }
         pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame +
@@ -584,6 +594,9 @@ static int tulip_copy_tx_buffers(TULIPState *s, struct tulip_descriptor *desc)
     int len2 = (desc->control >> TDES1_BUF2_SIZE_SHIFT) & TDES1_BUF2_SIZE_MASK;
 
     if (s->tx_frame_len + len1 > sizeof(s->tx_frame)) {
+        qemu_log_mask(LOG_GUEST_ERROR,
+                      "%s: descriptor overflow (ofs: %u, len:%d, size:%zu)\n",
+                      __func__, s->tx_frame_len, len1, sizeof(s->tx_frame));
         return -1;
     }
     if (len1) {
@@ -593,6 +606,9 @@ static int tulip_copy_tx_buffers(TULIPState *s, struct tulip_descriptor *desc)
     }
 
     if (s->tx_frame_len + len2 > sizeof(s->tx_frame)) {
+        qemu_log_mask(LOG_GUEST_ERROR,
+                      "%s: descriptor overflow (ofs: %u, len:%d, size:%zu)\n",
+                      __func__, s->tx_frame_len, len2, sizeof(s->tx_frame));
         return -1;
     }
     if (len2) {
Signature · 1 line
2.21.1
#3 lore raw

RFC PATCH 3/3 hw/net/tulip: Set descriptor error bit when lenght is incorrect

In reply to Philippe Mathieu-Daudé

When a frame lenght is incorrect, set the RDES0 'Error Summary' and 'Frame too long' bits. Then stop the receive process and trigger an abnormal interrupt. See [4.3.5 Receive Process].

Cc Li Qiang <[email protected]>
Cc Li Qiang <[email protected]>
Cc Ziming Zhang <[email protected]>
Cc Jason Wang <[email protected]>
Cc Prasad J Pandit <[email protected]>
Fixes 8ffb7265af ("check frame size and r/w data length")
---
 hw/net/tulip.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)
hw/net/tulip.c +12 -4
index 470f635acb..671f79b6f4 100644
@@ -158,7 +158,7 @@ static void tulip_next_rx_descriptor(TULIPState *s,
     s->current_rx_desc &= ~3ULL;
 }
 
-static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
+static int tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
 {
     int len1 = (desc->control >> RDES1_BUF1_SIZE_SHIFT) & RDES1_BUF1_SIZE_MASK;
     int len2 = (desc->control >> RDES1_BUF2_SIZE_SHIFT) & RDES1_BUF2_SIZE_MASK;
@@ -177,7 +177,8 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                           "(ofs: %u, len:%d, size:%zu)\n",
                           __func__, s->rx_frame_len, len,
                           sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
         }
         pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame +
             (s->rx_frame_size - s->rx_frame_len), len);
@@ -197,12 +198,15 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                           "(ofs: %u, len:%d, size:%zu)\n",
                           __func__, s->rx_frame_len, len,
                           sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
         }
         pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame +
             (s->rx_frame_size - s->rx_frame_len), len);
         s->rx_frame_len -= len;
     }
+
+    return 0;
 }
 
 static bool tulip_filter_address(TULIPState *s, const uint8_t *addr)
@@ -274,7 +278,11 @@ static ssize_t tulip_receive(TULIPState *s, const uint8_t *buf, size_t size)
             s->rx_frame_len = s->rx_frame_size;
         }
 
-        tulip_copy_rx_bytes(s, &desc);
+        if (tulip_copy_rx_bytes(s, &desc)) {
+            desc.status |= RDES0_ES | RDES0_TL; /* Error: frame too long */
+            s->csr[5] |= CSR5_RPS; /* Receive process stopped */
+            tulip_update_int(s);
+        }
 
         if (!s->rx_frame_len) {
             desc.status |= s->rx_status;
Signature · 1 line
2.21.1
#4 lore raw

Re: [RFC PATCH 3/3] hw/net/tulip: Set descriptor error bit when lenght is incorrect

Quoted 4 lines · Philippe Mathieu-Daudé

When a frame lenght is incorrect, set the RDES0 'Error Summary' and 'Frame too long' bits. Then stop the receive process and trigger an abnormal interrupt. See [4.3.5 Receive Process].

Quoted 9 lines
Cc Li Qiang <[email protected]>
Cc Li Qiang <[email protected]>
Cc Ziming Zhang <[email protected]>
Cc Jason Wang <[email protected]>
Cc Prasad J Pandit <[email protected]>
Fixes 8ffb7265af ("check frame size and r/w data length")

Hi Philippe:

It's still unclear to me that how this fixes the stuck. Did you mean guest trigger the error condition and then recvoer by abnormal interrupt?

If yes, this sounds still like a bug somewhere in the code.

Thanks

Quoted 57 lines
---
  hw/net/tulip.c | 16 ++++++++++++----
  1 file changed, 12 insertions(+), 4 deletions(-)
hw/net/tulip.c +12 -4
index 470f635acb..671f79b6f4 100644
@@ -158,7 +158,7 @@ static void tulip_next_rx_descriptor(TULIPState *s,
      s->current_rx_desc &= ~3ULL;
  }
 
-static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
+static int tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
  {
      int len1 = (desc->control >> RDES1_BUF1_SIZE_SHIFT) & RDES1_BUF1_SIZE_MASK;
      int len2 = (desc->control >> RDES1_BUF2_SIZE_SHIFT) & RDES1_BUF2_SIZE_MASK;
@@ -177,7 +177,8 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                            "(ofs: %u, len:%d, size:%zu)\n",
                            __func__, s->rx_frame_len, len,
                            sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
          }
          pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame +
              (s->rx_frame_size - s->rx_frame_len), len);
@@ -197,12 +198,15 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                            "(ofs: %u, len:%d, size:%zu)\n",
                            __func__, s->rx_frame_len, len,
                            sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
          }
          pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame +
              (s->rx_frame_size - s->rx_frame_len), len);
          s->rx_frame_len -= len;
      }
+
+    return 0;
  }
 
  static bool tulip_filter_address(TULIPState *s, const uint8_t *addr)
@@ -274,7 +278,11 @@ static ssize_t tulip_receive(TULIPState *s, const uint8_t *buf, size_t size)
              s->rx_frame_len = s->rx_frame_size;
          }
 
-        tulip_copy_rx_bytes(s, &desc);
+        if (tulip_copy_rx_bytes(s, &desc)) {
+            desc.status |= RDES0_ES | RDES0_TL; /* Error: frame too long */
+            s->csr[5] |= CSR5_RPS; /* Receive process stopped */
+            tulip_update_int(s);
+        }
 
          if (!s->rx_frame_len) {
              desc.status |= s->rx_status;
#5 lore raw

Re: [RFC PATCH 3/3] hw/net/tulip: Set descriptor error bit when lenght is incorrect

Quoted 17 lines · Jason Wang
Quoted 4 lines · Philippe Mathieu-Daudé

When a frame lenght is incorrect, set the RDES0 'Error Summary' and 'Frame too long' bits. Then stop the receive process and trigger an abnormal interrupt. See [4.3.5 Receive Process].

Quoted 9 lines
Cc Li Qiang <[email protected]>
Cc Li Qiang <[email protected]>
Cc Ziming Zhang <[email protected]>
Cc Jason Wang <[email protected]>
Cc Prasad J Pandit <[email protected]>
Fixes 8ffb7265af ("check frame size and r/w data length")
Quoted 5 lines

Hi Philippe:

It's still unclear to me that how this fixes the stuck. Did you mean guest trigger the error condition and then recvoer by abnormal interrupt?

If yes, this sounds still like a bug somewhere in the code.

Sadly the patches do not fix the bug.

Helge

Quoted 55 lines
Quoted 54 lines

---
  hw/net/tulip.c | 16 ++++++++++++----

  1 file changed, 12 insertions(+), 4 deletions(-)
hw/net/tulip.c +3 -1
index 470f635acb..671f79b6f4 100644
@@ -158,7 +158,7 @@ static void tulip_next_rx_descriptor(TULIPState *s,
       s->current_rx_desc &= ~3ULL;
   }
   -static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
+static int tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
   {
       int len1 = (desc->control >> RDES1_BUF1_SIZE_SHIFT) & RDES1_BUF1_SIZE_MASK;
       int len2 = (desc->control >> RDES1_BUF2_SIZE_SHIFT) & RDES1_BUF2_SIZE_MASK;
@@ -177,7 +177,8 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                             "(ofs: %u, len:%d, size:%zu)\n",
                             __func__, s->rx_frame_len, len,
                             sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
           }

          pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame + (s->rx_frame_size - s->rx_frame_len), len);

hw/net/tulip.c +2 -1
@@ -197,12 +198,15 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
                             "(ofs: %u, len:%d, size:%zu)\n",
                             __func__, s->rx_frame_len, len,
                             sizeof(s->rx_frame));
-            return;
+            s->rx_frame_len = 0;
+            return -1;
           }

          pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame + (s->rx_frame_size - s->rx_frame_len), len); s->rx_frame_len -= len;
      }
+
+    return 0;
  }
    static bool tulip_filter_address(TULIPState *s, const uint8_t *addr)

hw/net/tulip.c +5 -0
@@ -274,7 +278,11 @@ static ssize_t tulip_receive(TULIPState *s, const uint8_t *buf, size_t size)
               s->rx_frame_len = s->rx_frame_size;
           }
   -        tulip_copy_rx_bytes(s, &desc);
+        if (tulip_copy_rx_bytes(s, &desc)) {
+            desc.status |= RDES0_ES | RDES0_TL; /* Error: frame too long */
+            s->csr[5] |= CSR5_RPS; /* Receive process stopped */
+            tulip_update_int(s);
+        }
             if (!s->rx_frame_len) {
               desc.status |= s->rx_status;
#6 lore raw

Re: [RFC PATCH 3/3] hw/net/tulip: Set descriptor error bit when lenght is incorrect

In reply to Philippe Mathieu-Daudé
Quoted 1 line · Philippe Mathieu-Daudé

When a frame lenght is incorrect, set the RDES0 'Error Summary'

length (here, and in the subject)

Quoted 3 lines

and 'Frame too long' bits. Then stop the receive process and trigger an abnormal interrupt. See [4.3.5 Receive Process].

Signature · 3 lines
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org
#7 lore raw
In reply to Philippe Mathieu-Daudé

* Philippe Mathieu-Daudé <[email protected]>:

Quoted 8 lines

Attempt to fix the launchpad bug filled by Helge:

In a qemu-system-hppa system, qemu release v5.0.0-rc, the tulip nic driver is broken. The tulip nic is detected, even getting DHCP info does work. But when trying to download bigger files via network, the tulip driver gets stuck.

Quoted 8 lines

Philippe Mathieu-Daudé (3):
hw/net/tulip: Fix 'Descriptor Error' definition hw/net/tulip: Log descriptor overflows
hw/net/tulip: Set descriptor error bit when lenght is incorrect

 hw/net/tulip.h |  2 +-
 hw/net/tulip.c | 32 ++++++++++++++++++++++++++++----
 2 files changed, 29 insertions(+), 5 deletions(-)

Philippe, thanks for your efforts. Sadly your patch did not fixed the bug itself, but it had some nice cleanups which should be included at some point.

Regarding the tulip hang reported by me, the patch below does fix the issue.

[PATCH] Fix tulip rx hang

Cc Prasad J Pandit <[email protected]>
Fixes 8ffb7265af ("check frame size and r/w data length")

Commit 8ffb7265af ("check frame size and r/w data length") introduced checks to prevent accesses outside of the rx/tx buffers. But the new checks were plain wrong. rx_frame_len does count backwards, and the surrounding code ensures that rx_frame_len will not be bigger than rx_frame_size. Remove those checks again.

hw/net/tulip.c +0 -6
index 1295f51d07..59d21defcc 100644
@@ -171,9 +171,6 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
             len = s->rx_frame_len;
         }
 
-        if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
-            return;
-        }
         pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame +
             (s->rx_frame_size - s->rx_frame_len), len);
         s->rx_frame_len -= len;
@@ -186,9 +183,6 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
             len = s->rx_frame_len;
         }
 
-        if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
-            return;
-        }
         pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame +
             (s->rx_frame_size - s->rx_frame_len), len);
         s->rx_frame_len -= len;
#8 lore raw
Quoted 52 lines · Helge Deller

* Philippe Mathieu-Daudé <[email protected]>:

Quoted 8 lines

Attempt to fix the launchpad bug filled by Helge:

In a qemu-system-hppa system, qemu release v5.0.0-rc, the tulip nic driver is broken. The tulip nic is detected, even getting DHCP info does work. But when trying to download bigger files via network, the tulip driver gets stuck.

Quoted 8 lines

Philippe Mathieu-Daudé (3):
hw/net/tulip: Fix 'Descriptor Error' definition hw/net/tulip: Log descriptor overflows
hw/net/tulip: Set descriptor error bit when lenght is incorrect

  hw/net/tulip.h |  2 +-
  hw/net/tulip.c | 32 ++++++++++++++++++++++++++++----
  2 files changed, 29 insertions(+), 5 deletions(-)

Philippe, thanks for your efforts. Sadly your patch did not fixed the bug itself, but it had some nice cleanups which should be included at some point.

Regarding the tulip hang reported by me, the patch below does fix the issue.

[PATCH] Fix tulip rx hang

Cc Prasad J Pandit <[email protected]>
Fixes 8ffb7265af ("check frame size and r/w data length")

Commit 8ffb7265af ("check frame size and r/w data length") introduced checks to prevent accesses outside of the rx/tx buffers. But the new checks were plain wrong. rx_frame_len does count backwards, and the surrounding code ensures that rx_frame_len will not be bigger than rx_frame_size. Remove those checks again.

hw/net/tulip.c +0 -3
index 1295f51d07..59d21defcc 100644
@@ -171,9 +171,6 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
              len = s->rx_frame_len;
          }
 
-        if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
-            return;
-        }
          pci_dma_write(&s->dev, desc->buf_addr1, s->rx_frame +
              (s->rx_frame_size - s->rx_frame_len), len);
          s->rx_frame_len -= len;

hw/net/tulip.c static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc) · line 188

Quoted 9 lines
hw/net/tulip.c +0 -3
@@ -186,9 +183,6 @@ static void tulip_copy_rx_bytes(TULIPState *s, struct tulip_descriptor *desc)
              len = s->rx_frame_len;
          }
 
-        if (s->rx_frame_len + len > sizeof(s->rx_frame)) {
-            return;
-        }
          pci_dma_write(&s->dev, desc->buf_addr2, s->rx_frame +
              (s->rx_frame_size - s->rx_frame_len), len);
          s->rx_frame_len -= len;

Looks good to me.

Would you please send a formal patch and cc Peter.

Consider we are about to release 5.0, it's better for him to apply the patch directly.

Thanks

#9 lore raw
Quoted 6 lines · Jason Wang

Looks good to me.

Would you please send a formal patch and cc Peter.

Consider we are about to release 5.0, it's better for him to apply the patch directly.

I am not applying any further patches for 5.0 unless they come with an attached rock-solid justification for why we should delay the release again for them.

thanks
-- PMM

#10 lore raw
Quoted 6 lines · Peter Maydell
Quoted 4 lines · Jason Wang

Looks good to me.

Would you please send a formal patch and cc Peter.

Quoted 8 lines
Quoted 2 lines

Consider we are about to release 5.0, it's better for him to apply the patch directly.

I am not applying any further patches for 5.0 unless they come with an attached rock-solid justification for why we should delay the release again for them.

thanks
-- PMM

Ok.

I will queue that patch for 5.1.

Thanks

Quoted 1 line
#11 lore raw

Hi Jason,

Quoted 13 lines · Jason Wang
Quoted 12 lines · Peter Maydell
Quoted 6 lines · Jason Wang

Looks good to me.

Would you please send a formal patch and cc Peter.

Consider we are about to release 5.0, it's better for him to apply the patch directly.

I am not applying any further patches for 5.0 unless they come with an attached rock-solid justification for why we should delay the release again for them.

Quoted 7 lines
Quoted 2 lines

thanks
-- PMM

Ok.

I will queue that patch for 5.1.

Can you queue patches #1 and #2?

Quoted 7 lines

Thanks

Quoted 1 line
#12 lore raw
Quoted 20 lines · Philippe Mathieu-Daudé

Hi Jason,

Quoted 17 lines · Jason Wang
Quoted 14 lines · Peter Maydell
Quoted 6 lines · Jason Wang

Looks good to me.

Would you please send a formal patch and cc Peter.

Consider we are about to release 5.0, it's better for him to apply the patch directly.

I am not applying any further patches for 5.0 unless they come with an attached rock-solid justification for why we should delay the release again for them.

thanks
-- PMM

Quoted 5 lines
Quoted 3 lines

Ok.

I will queue that patch for 5.1.

Can you queue patches #1 and #2?

Queued.

Thanks

Quoted 9 lines
Quoted 8 lines

Thanks

Quoted 1 line